Updated hourly
LLM comparison for builders and product teams
Last successful sync: October 18, 2026 at 07:00 AM UTC
Previous snapshot: October 17, 2026
How AI ratings work
Scores estimate overall capability: reasoning 35%, coding 30%, instruction following 25%, and reliability 10%. AI reviews use published sources and are refreshed monthly. The top three are the highest-rated distinct models with current, medium- or high-confidence reviews. Scores are AI assessments, not hands-on benchmark results. Price does not affect the ranking.
Top 3 models by AI rating
#1OpenAI: GPT-5.5 Pro
OpenAI
OpenAI’s docs describe GPT-5.5 Pro as a tougher-problem variant that uses more compute to produce smarter, more precise responses, with broad tool support and 1M context. That supports frontier-level capability, though public evidence still comes mainly from OpenAI’s own evaluations and documentation.
Input
$30.00
Context
1.1M
#2Anthropic: Claude Opus 5.5
Anthropic
Anthropic’s system-card hub and launch materials support Opus 5.5 as a frontier-class model, with especially strong coding and agentic performance. The main limitation is that public evidence is still mostly vendor-run, so its exact position relative to the newest rivals remains somewhat uncertain.
Input
$4.000
Context
1M
#3OpenAI: GPT-5.5
OpenAI
OpenAI describes GPT-5.5 as a frontier model for complex professional work, with the latest guidance recommending it for production API usage. The evidence supports very strong general capability, though exact rank versus GPT-5.5 Pro and other frontier rivals remains partly vendor-defined.
Input
$5.000
Context
1.1M
444 models shown
| Model | Provider | AI rating | Input / 1M | Output / 1M | Context | Capabilities / strengths | Best for |
|---|---|---|---|---|---|---|---|
| OpenAI: GPT-5.5 Pro Listed Apr 24, 2026 OpenAI’s docs describe GPT-5.5 Pro as a tougher-problem variant that uses more compute to produce smarter, more precise responses, with broad tool support and 1M context. That supports frontier-level capability, though public evidence still comes mainly from OpenAI’s own evaluations and documentation. | OpenAI | 9.2/10 | $30.00 | $180.00 | 1.1M | Stronger-compute variant aimed at harder problems, Excellent broad general-purpose capability, Very strong tool support and long context, Good for high-stakes, accuracy-sensitive workloads | Complex reasoning and analysis, High-stakes assistant workflows, Tool-heavy agents, Long-context professional tasks |
| OpenAI: GPT-5.5 Pro (batch) Listed Apr 24, 2026 OpenAI’s docs describe GPT-5.5 Pro as a tougher-problem variant that uses more compute to produce smarter, more precise responses, with broad tool support and 1M context. That supports frontier-level capability, though public evidence still comes mainly from OpenAI’s own evaluations and documentation. | OpenAI | 9.2/10 | $15.00 | $90.00 | 1.1M | Stronger-compute variant aimed at harder problems, Excellent broad general-purpose capability, Very strong tool support and long context, Good for high-stakes, accuracy-sensitive workloads | Complex reasoning and analysis, High-stakes assistant workflows, Tool-heavy agents, Long-context professional tasks |
| Anthropic: Claude Opus 5.5 Listed Sep 22, 2026 Anthropic’s system-card hub and launch materials support Opus 5.5 as a frontier-class model, with especially strong coding and agentic performance. The main limitation is that public evidence is still mostly vendor-run, so its exact position relative to the newest rivals remains somewhat uncertain. | Anthropic | 9.1/10 | $4.000 | $20.00 | 1M | Top-tier coding and agentic-work focus, Strong general reasoning for demanding professional tasks, Current Anthropic materials place it at the top of the Opus tier, Good fit for defensive cyber and complex enterprise workflows | Large codebase changes and debugging, Long-horizon agentic work, Enterprise knowledge-work assistance, Security-focused workflows |
| Anthropic: Claude Opus 5.5 (batch) Listed Sep 22, 2026 Anthropic’s system-card hub and launch materials support Opus 5.5 as a frontier-class model, with especially strong coding and agentic performance. The main limitation is that public evidence is still mostly vendor-run, so its exact position relative to the newest rivals remains somewhat uncertain. | Anthropic | 9.1/10 | $2.000 | $10.00 | 1M | Top-tier coding and agentic-work focus, Strong general reasoning for demanding professional tasks, Current Anthropic materials place it at the top of the Opus tier, Good fit for defensive cyber and complex enterprise workflows | Large codebase changes and debugging, Long-horizon agentic work, Enterprise knowledge-work assistance, Security-focused workflows |
| OpenAI: GPT-5.5 Listed Apr 24, 2026 OpenAI describes GPT-5.5 as a frontier model for complex professional work, with the latest guidance recommending it for production API usage. The evidence supports very strong general capability, though exact rank versus GPT-5.5 Pro and other frontier rivals remains partly vendor-defined. | OpenAI | 9.1/10 | $5.000 | $30.00 | 1.1M | Frontier model for complex professional work, Strong tool use and long-context workflows, Improved baseline over GPT-5.4 for difficult tasks, Well-suited to mixed reasoning, writing, and coding | General-purpose professional assistance, Long-context project work, Tool-heavy task execution, Reasoning across mixed document and code tasks |
| OpenAI: GPT-5.5 (batch) Listed Apr 24, 2026 OpenAI describes GPT-5.5 as a frontier model for complex professional work, with the latest guidance recommending it for production API usage. The evidence supports very strong general capability, though exact rank versus GPT-5.5 Pro and other frontier rivals remains partly vendor-defined. | OpenAI | 9.1/10 | $2.500 | $15.00 | 1.1M | Frontier model for complex professional work, Strong tool use and long-context workflows, Improved baseline over GPT-5.4 for difficult tasks, Well-suited to mixed reasoning, writing, and coding | General-purpose professional assistance, Long-context project work, Tool-heavy task execution, Reasoning across mixed document and code tasks |
| Google: Gemini 4 Argon Listed Oct 13, 2026 Google’s current model card presents Argon as a true frontier model with broad multimodal capability and 1M context; however, the evidence is mainly vendor-run, so exact rank versus other frontier models remains uncertain. It looks excellent for hard, multi-step work, with the main caution being that public comparative evidence is still thin. | 9.0/10 | $2.000 | $10.00 | 1M | Frontier-scale reasoning across coding, math, science, and enterprise work, Native multimodal input including text, image, audio, and video, Large 1M-token context for long-horizon workflows, Strong fit for deep investigations and complex engineering | Complex software engineering and refactoring, Scientific or mathematical analysis, Multimodal document understanding, Long multi-step agent workflows | |
| Google: Gemini 4 Argon (batch) Listed Oct 13, 2026 Google’s current model card presents Argon as a true frontier model with broad multimodal capability and 1M context; however, the evidence is mainly vendor-run, so exact rank versus other frontier models remains uncertain. It looks excellent for hard, multi-step work, with the main caution being that public comparative evidence is still thin. | 9.0/10 | $1.000 | $5.000 | 1M | Frontier-scale reasoning across coding, math, science, and enterprise work, Native multimodal input including text, image, audio, and video, Large 1M-token context for long-horizon workflows, Strong fit for deep investigations and complex engineering | Complex software engineering and refactoring, Scientific or mathematical analysis, Multimodal document understanding, Long multi-step agent workflows | |
| OpenAI: GPT-6 Sol Pro Listed Sep 22, 2026 OpenAI describes GPT-6 Sol as the flagship family member and says Sol Pro is the same underlying model with pro reasoning mode. Public benchmark detail is limited here, but the available primary sources support a frontier-level estimate. | OpenAI | 9.0/10 AI review detailsLow confidence · Sep 29, 2026 Coding: 9/10 Reasoning: 9.2/10 Reliability: 8.4/10 Instructions: 8.8/10 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $2.000 | $10.00 | 1.1M | Flagship GPT-6 capability tier, Strong coding and agentic-workflow orientation, Excellent tool use and structured outputs, Vision/image input support | Hard reasoning and research tasks, Advanced coding and software engineering help, Agentic tool-using workflows, Multimodal problem solving |
| OpenAI: GPT-6 Sol Pro (batch) Listed Sep 22, 2026 This batch alias should not materially change model capability relative to GPT-6 Sol Pro. Because public evidence on the exact batch variant is sparse, I’m less certain than for the non-batch model. | OpenAI | 9.0/10 | $1.000 | $5.000 | 1.1M | Same underlying model as GPT-6 Sol Pro, Batch-oriented deployment, Strong coding and reasoning profile, Tool and structured-output support | Batch frontier reasoning jobs, Large-scale coding assistance, Queued agent pipelines, Offline multimodal analysis |
| OpenAI: GPT-5.4 Listed Mar 5, 2026 OpenAI positions GPT-5.4 as a flagship model for complex professional work, and the docs show broad tool support plus 1M-context capability. That indicates frontier-tier capability, but the public evidence emphasizes product fit more than independent comparative leadership. | OpenAI | 9.0/10 | $2.500 | $15.00 | 1.1M | Flagship general-purpose model for professional work, Strong coding and workflow automation support, Large 1.05M context window, Broad tool ecosystem including computer use and structured outputs | Professional analysis and synthesis, Production software development, Agentic workflow automation, Long-context retrieval and planning |
| OpenAI: GPT-5.4 (batch) Listed Mar 5, 2026 OpenAI positions GPT-5.4 as a flagship model for complex professional work, and the docs show broad tool support plus 1M-context capability. That indicates frontier-tier capability, but the public evidence emphasizes product fit more than independent comparative leadership. | OpenAI | 9.0/10 | $1.250 | $7.500 | 1.1M | Flagship general-purpose model for professional work, Strong coding and workflow automation support, Large 1.05M context window, Broad tool ecosystem including computer use and structured outputs | Professional analysis and synthesis, Production software development, Agentic workflow automation, Long-context retrieval and planning |
| OpenAI: GPT-5 Mini (batch) Listed Aug 7, 2025 Primary OpenAI docs describe GPT-5 mini as a faster, more cost-efficient GPT-5 variant with 400K context, image input, and reasoning token support. Evidence suggests strong all-around capability, with the main limitation being that it is the smaller, lower-cost sibling rather than the flagship. | OpenAI | 9.0/10 | $0.125 | $1.000 | 400k | Very strong general reasoning for a compact model, Good coding and agentic task performance, Strong instruction adherence and steerability, Supports images and large-context workflows | High-volume structured extraction and transformation, Coding assistants for routine-to-hard tasks, Tool-using workflows with tight latency budgets, Multimodal Q&A and document/image analysis |
| Google: Gemini 3.1 Pro Preview Custom Tools Listed Feb 25, 2026 This appears to be a tool-optimized variant of Gemini 3.1 Pro Preview rather than a different base model. The tool-use advantage is plausible, but there is not enough public measurement to score it precisely. | 8.7/10 | $2.000 | $12.00 | 1M | Tool-selection improvements, Strong multimodal base model, Good reasoning controls, Useful for tool-heavy agents | Tool-heavy assistants, Workflow orchestration, Multimodal tasks, Agent experiments | |
| OpenAI: GPT-5.1 Listed Nov 13, 2025 OpenAI’s materials describe GPT-5.1 as the flagship model for coding and agentic tasks, with configurable reasoning effort and a substantial context window. It is very capable, but the later GPT-5.4 and GPT-5.5 releases appear to be the stronger current mainline options. | OpenAI | 8.7/10 | $1.250 | $10.00 | 400k | Flagship coding/agentic model in its generation, Configurable reasoning effort, Strong instruction adherence and conversational quality, Good general-purpose performance with broad tool support | Coding and code review, Agentic workflows, General reasoning and writing, Tool-using assistants |
| OpenAI: GPT-5.1 (batch) Listed Nov 13, 2025 OpenAI’s materials describe GPT-5.1 as the flagship model for coding and agentic tasks, with configurable reasoning effort and a substantial context window. It is very capable, but the later GPT-5.4 and GPT-5.5 releases appear to be the stronger current mainline options. | OpenAI | 8.7/10 | $0.625 | $5.000 | 400k | Flagship coding/agentic model in its generation, Configurable reasoning effort, Strong instruction adherence and conversational quality, Good general-purpose performance with broad tool support | Coding and code review, Agentic workflows, General reasoning and writing, Tool-using assistants |
| OpenAI: GPT-6.1 Sol Pro Listed Sep 29, 2026 This wrapper appears to point at GPT-6 Sol plus a higher-effort pro mode, but the exact OpenRouter name is not independently documented by OpenAI. Scores reflect a cautious inference from the official GPT-6 Sol evidence, not a verified separate model class. | OpenAI | 8.6/10 | $2.000 | $10.00 | 1.1M | likely strong agentic coding, broad tool support, good for computer use, pro mode may improve complex tasks | agentic coding, computer-use workflows, document-heavy professional work, mixed tool-using assistants |
| OpenAI: GPT-6.1 Sol Pro (batch) Listed Sep 29, 2026 This wrapper appears to point at GPT-6 Sol plus a higher-effort pro mode, but the exact OpenRouter name is not independently documented by OpenAI. Scores reflect a cautious inference from the official GPT-6 Sol evidence, not a verified separate model class. | OpenAI | 8.6/10 | $1.000 | $5.000 | 1.1M | likely strong agentic coding, broad tool support, good for computer use, pro mode may improve complex tasks | agentic coding, computer-use workflows, document-heavy professional work, mixed tool-using assistants |
| SpaceXAI: Grok 4.7 Listed Sep 21, 2026 Grok 4.7 appears to be xAI’s newest flagship and should be strong on coding and agentic tasks. The evidence is mostly vendor-led, so confidence is lower than for the most established systems. | xAI | 8.6/10 | $2.000 | $6.000 | 500k | strong coding focus, good agentic tasks, self-verification emphasis, competitive price-performance | software engineering, long-running coding agents, knowledge work, tool-based automation |
| OpenAI: GPT-5.3-Codex Listed Feb 24, 2026 OpenAI explicitly calls GPT-5.3-Codex its most capable agentic coding model to date, so coding gets a higher score than broad general-purpose reasoning. Because it is specialized, it is less balanced than the mainline frontier models for non-coding work. | OpenAI | 8.6/10 | $1.750 | $14.00 | 400k | Very strong agentic coding specialization, Optimized for Codex-like environments, Long-running coding workflows, High-end coding effort settings | Large-scale software engineering, Multi-step coding agents, Code modification and review, Long-horizon dev tasks |
| Anthropic: Claude Sonnet 4.6 Listed Feb 17, 2026 Anthropic’s system-card listing and related materials place Sonnet 4.6 as a strong frontier-class model, but clearly below the newest Opus tier in this batch. It should be excellent for everyday professional work, though not the very top choice for maximum capability. | Anthropic | 8.6/10 | $3.000 | $15.00 | 1M | Strong frontier-class performance in practical work, Good coding and codebase-navigation ability, Solid multimodal and agentic support, Balanced quality/cost among top models | Iterative software development, Code review and codebase navigation, General professional assistance, Multimodal analysis |
| Anthropic: Claude Sonnet 4.6 (batch) Listed Feb 17, 2026 Anthropic’s system-card listing and related materials place Sonnet 4.6 as a strong frontier-class model, but clearly below the newest Opus tier in this batch. It should be excellent for everyday professional work, though not the very top choice for maximum capability. | Anthropic | 8.6/10 | $1.500 | $7.500 | 1M | Strong frontier-class performance in practical work, Good coding and codebase-navigation ability, Solid multimodal and agentic support, Balanced quality/cost among top models | Iterative software development, Code review and codebase navigation, General professional assistance, Multimodal analysis |
| OpenAI: GPT-5.1-Codex-Max Listed Dec 4, 2025 OpenAI’s system card presents GPT-5.1-Codex-Max as a frontier agentic coding model trained on software engineering and long-context compaction workflows. That supports excellent coding scores, but the specialization makes it less balanced than a flagship general-purpose model. | OpenAI | 8.6/10 | $1.250 | $10.00 | 400k | Top-tier long-running agentic coding specialization, Compaction for multi-window software tasks, Built for real-world software engineering workflows, Very strong code-change and code-review fit | Long-horizon software development, Large codebase refactoring, PR creation and review, Agentic coding in compaction-heavy workflows |
| OpenAI: o3 Pro Listed Jun 10, 2025 OpenAI positions o3-pro as a more-compute version of o3 for better responses on hard problems, which supports a high reasoning score. It is still an older reasoning line, so I would place it below the newer GPT-5.x flagships in overall balance. | OpenAI | 8.5/10 | $20.00 | $80.00 | 200k | More-compute reasoning variant for hard problems, Strong consistency on difficult tasks, Good for deliberative workflows, Supports core OpenAI tool features | Hard reasoning problems, Careful analysis and planning, Legacy o-series workflows, Tool-using assistant backends |
| OpenAI: GPT-6.1 Sol Listed Sep 29, 2026 The exact name is not an official OpenAI model label in the sources found, so this is an inference from the OpenRouter wrapper to GPT-6 Sol. Capability appears strong, but the version mapping is uncertain. | OpenAI | 8.4/10 | $2.000 | $10.00 | 1.1M | strong coding and agentic workflow fit, good tool support, broad multimodal utility, likely efficient for everyday professional work | software development, tool-using assistants, document analysis, computer-use workflows |
| OpenAI: GPT-6.1 Sol (batch) Listed Sep 29, 2026 The exact name is not an official OpenAI model label in the sources found, so this is an inference from the OpenRouter wrapper to GPT-6 Sol. Capability appears strong, but the version mapping is uncertain. | OpenAI | 8.4/10 | $1.000 | $5.000 | 1.1M | strong coding and agentic workflow fit, good tool support, broad multimodal utility, likely efficient for everyday professional work | software development, tool-using assistants, document analysis, computer-use workflows |
| Qwen: Qwen3.8 Max (0902) Listed Sep 3, 2026 Evidence supports a flagship-class Qwen model, but independent comparison data is thinner than for the best-known closed frontier systems. I score it conservatively because public validation is limited. | Qwen | 8.4/10 | $2.000 | $6.000 | 1M | Flagship-class Qwen model, Strong multimodal support, Good coding and agent use, Large context window | Multimodal reasoning, Coding assistance, Agentic workflows, High-volume document tasks |
| Anthropic: Claude Opus 4.5 Listed Nov 24, 2025 Opus 4.5 was a leading model at release, but Anthropic’s newer Opus 5.5 and related successors are stronger, so it no longer occupies the top spot here. It remains very capable, just a generation behind the newest frontier models. | Anthropic | 8.4/10 | $5.000 | $25.00 | 200k | Still a strong frontier-class Opus model, Good coding and agentic workflow performance, Broad multimodal and professional-use support, Useful for long-horizon computer-use tasks | Software engineering assistance, Agentic workflows, Multimodal analysis, Professional knowledge work |
| Anthropic: Claude Opus 4.5 (batch) Listed Nov 24, 2025 Opus 4.5 was a leading model at release, but Anthropic’s newer Opus 5.5 and related successors are stronger, so it no longer occupies the top spot here. It remains very capable, just a generation behind the newest frontier models. | Anthropic | 8.4/10 | $2.500 | $12.50 | 200k | Still a strong frontier-class Opus model, Good coding and agentic workflow performance, Broad multimodal and professional-use support, Useful for long-horizon computer-use tasks | Software engineering assistance, Agentic workflows, Multimodal analysis, Professional knowledge work |
| Qwen: Qwen3.8 2.4T A95B Listed Aug 12, 2026 The open-weight Qwen3.8 variant appears close to flagship class within its family. Independent comparison data is limited, so the score is conservative. | Qwen | 8.3/10 | $2.000 | $6.000 | 1M | Open-weight MoE flagship variant, Large context window, Good tool use, Strong multimodal orientation | Open-weight research, Coding workflows, Agent prototypes, Long-context tasks |
| SpaceXAI: Grok 4.6 Listed Aug 12, 2026 Grok 4.6 looks like a strong frontier model with clear coding and agentic strengths. The evidence is largely vendor-led, so I keep confidence a bit below the most established closed systems. | xAI | 8.3/10 | $2.000 | $6.000 | 500k | strong frontier coding, good knowledge work, agentic task support, competitive general utility | coding assistance, multi-step automation, research workflows, technical Q&A |
| Google: Gemini 3.8 Flash Listed Sep 2, 2026 Google describes Gemini 3.8 Flash as its most intelligent Flash model, with strong software-engineering and agentic performance plus 1M context. That makes it very capable, but the Flash positioning still suggests it sits below the top reasoning-oriented frontier models in this batch. | 8.2/10 | $0.825 | $4.125 | 1M | Fast Flash-tier model with strong capability, 1M-token context window, Good agentic and software-engineering performance for its class, Strong built-in tool and multimodal support | High-throughput coding assistants, Agentic workflows with latency constraints, Enterprise document workflows, Multimodal utility tasks | |
| Google: Gemini 3.8 Flash (batch) Listed Sep 2, 2026 Google describes Gemini 3.8 Flash as its most intelligent Flash model, with strong software-engineering and agentic performance plus 1M context. That makes it very capable, but the Flash positioning still suggests it sits below the top reasoning-oriented frontier models in this batch. | 8.2/10 | $0.375 | $1.875 | 1M | Fast Flash-tier model with strong capability, 1M-token context window, Good agentic and software-engineering performance for its class, Strong built-in tool and multimodal support | High-throughput coding assistants, Agentic workflows with latency constraints, Enterprise document workflows, Multimodal utility tasks | |
| OpenAI: GPT-5.4 Pro Listed Mar 5, 2026 GPT-5.4 Pro was a major OpenAI model, but newer GPT-5.6 and GPT-6 lines now sit above it. | OpenAI | 8.0/10 | $30.00 | $180.00 | 1.1M | Still a strong general-purpose OpenAI model., Good instruction following and code quality for its generation., Broad capability across tasks., Large context window. | General assistant tasks, Analysis and drafting, Coding help, Long-context work |
| OpenAI: GPT-5.4 Pro (batch) Listed Mar 5, 2026 GPT-5.4 Pro was a major OpenAI model, but newer GPT-5.6 and GPT-6 lines now sit above it. | OpenAI | 8.0/10 | $15.00 | $90.00 | 1.1M | Still a strong general-purpose OpenAI model., Good instruction following and code quality for its generation., Broad capability across tasks., Large context window. | General assistant tasks, Analysis and drafting, Coding help, Long-context work |
| MiniMax: MiniMax M2 Listed Oct 23, 2025 MiniMax describes M2 as an agent-focused coding model with planning and tool use. The evidence is promising but mostly vendor claims, so I’m confident in the direction but not the exact comparative rank. | MiniMax | 8.0/10 | $0.255 | $1.020 | 205k | Built for agents and code, Planning-oriented design, Good tool calling focus, End-to-end development workflow fit | Coding agents, Planning-heavy workflows, Tool-driven development, Automation prototypes |
| OpenAI: gpt-oss-120b (batch) Listed Aug 5, 2025 OpenAI’s docs and model card describe gpt-oss-120b as the strongest open-weight model in the family, built for agentic use and showing strong reasoning/coding results. | OpenAI | 8.0/10 | $0.030 | $0.136 | 131k | OpenAI positions it as the strongest open-weight model in the family., Docs report 117B total / 5.1B active parameters and Apache 2.0 licensing., Built for agentic use with function calling, browsing, Python, and structured outputs., Launch material cites strong reasoning and coding results. | Agentic assistants and tool-using workflows., Coding and software-adjacent tasks., Self-hosted production deployments., Reasoning-heavy enterprise automation. |
| OpenAI: o1 Listed Dec 17, 2024 OpenAI’s docs and system card show o1 as a prior full o-series reasoning model with strong reasoning behavior, but it is older than the GPT-5 family and deprecated. It remains capable, yet newer models in this batch are generally stronger and better supported. | OpenAI | 8.0/10 | $15.00 | $60.00 | 200k | Strong reasoning-focused architecture for its generation, Still useful for careful, deliberative tasks, Good historical benchmark of OpenAI’s reasoning line, Supports function calling and structured outputs | Deliberate reasoning tasks, Analytical Q&A, Legacy workflows that still use o1, Tasks that benefit from chain-of-thought style planning |
| NVIDIA: Nemotron 3.5 Super VL Listed Oct 15, 2026 Evidence suggests a strong multimodal agent model with especially good tool use, instruction following, and competitive coding. It is still a vendor-reported system with some claims based on internal or comparative evaluations, so I’d place it below frontier generalists. | NVIDIA | 7.9/10 | $0.100 vs prev -0.050 | $0.400 vs prev -0.100 | 1M | Strong multimodal agent workflow support across text, image, document, and video inputs., Good tool-use and instruction-following orientation, with benchmarked agent-task performance., Competitive coding and terminal-use performance, including SWE-bench and Terminal-Bench mentions., Long-context design and reasoning-budget controls suit complex, multi-step tasks. | Document and chart understanding with grounded extraction and summarization., Agentic workflows that combine tool use, reasoning, and visual inputs., Coding assistants for code review, debugging, and terminal-based tasks., Enterprise multimodal assistants for policies, reports, and workflow automation. |
| Pareto 26.10 Listed Oct 15, 2026 Evidence from Unbiased’s model card and blog suggests frontier-leaning general capability, especially in reasoning, coding, and agentic tasks. Confidence is limited by vendor-run benchmark methodology and preliminary comparison details, so reliability is solid but not proven best-in-class. | unbiased | 7.9/10 | $0.800 | $3.200 | 1M | Strong reported performance on GPQA, HLE, DeepSWE, and Terminal-Bench relative to many peers, Multimodal input support, Designed for research, coding, and agentic workflows, Large context window (1,048,576 tokens) on OpenRouter | Research-heavy assistant work, Agentic coding and repo-level debugging, Long-context document and multimodal analysis, Tool-using workflow orchestration |
| Meta: Muse Wonder Listed Oct 12, 2026 Meta’s own launch and OpenRouter’s model page both position Muse Wonder as a frontier agentic model with multimodal inputs, tool use, and a 1M-token context window. Capability appears strong, but independent benchmark evidence is still limited, so confidence is only medium. | Meta | 7.9/10 | $4.000 | $20.00 | 1M | Strong agentic task handling with planning, delegation, and error recovery, Multimodal input support: text, images, video, audio, and PDFs, Designed for long-context work with a 1M-token window, Supports tool calling and structured outputs | Complex multi-step personal or professional assistants, Multi-file coding and repository refactors, Long-document analysis across PDFs and mixed media, Browser- or app-based agent workflows |
| Sakana: Fugu Max Listed Sep 11, 2026 Sakana’s official release describes Fugu Max as a learned multi-agent orchestration system, not a monolithic model, so its capability depends on routing. That suggests strong workflow performance, but also more variance than a single frontier model. | Sakana | 7.9/10 | $2.000 | $6.000 | 1M | Orchestrates multiple models rather than relying on one backbone, Designed for agentic and cost-aware task routing, Supports text, images, files, tools, and structured output, Official release targets real-world workflow efficiency | Agentic multi-step work, Mixed text/image/document workflows, Cost-sensitive task routing across models |
| Qwen: Qwen3.8 Flash Listed Aug 26, 2026 Qwen’s Flash-Next materials suggest a capable sparse-MoE multimodal model built for agentic work, but the exact OpenRouter slug is ambiguous. I’m confident it is strong, but less confident about the precise version-to-version placement. | Qwen | 7.9/10 | $0.150 | $0.470 | 1M | Very cost-efficient, Strong multimodal reasoning design, Long-context suitability, Useful for agentic workflows | High-throughput assistant work, Document and codebase analysis, Cost-sensitive agent pipelines, Long-video or chart analysis |
| Z.ai: GLM 5.3 Listed Aug 18, 2026 GLM 5.3 looks strong on coding and long-horizon tasks, but the public evidence is mostly vendor-provided rather than independent. That supports a solid upper-mid frontier assessment, with more uncertainty than the leading OpenAI and Anthropic models. | Z.ai | 7.9/10 | $0.079 | $4.800 | 1M | Large-context reasoning model aimed at complex coding, Vendor sources indicate strong long-horizon agent performance, Broad text-generation capability, Competitive value relative to frontier peers | Software engineering assistance, Long-context document analysis, Agent workflows, High-volume reasoning tasks |
| Z.ai: GLM 5.3 (batch) Listed Aug 18, 2026 GLM 5.3 looks strong on coding and long-horizon tasks, but the public evidence is mostly vendor-provided rather than independent. That supports a solid upper-mid frontier assessment, with more uncertainty than the leading OpenAI and Anthropic models. | Z.ai | 7.9/10 | $0.450 | $2.000 | 1M | Large-context reasoning model aimed at complex coding, Vendor sources indicate strong long-horizon agent performance, Broad text-generation capability, Competitive value relative to frontier peers | Software engineering assistance, Long-context document analysis, Agent workflows, High-volume reasoning tasks |
| Meta: Muse Glimmer 30B Listed Aug 9, 2026 Meta’s current materials present Muse Glimmer as a 30B open-weights multimodal model tuned for tool use, long tasks, and failure recovery, with benchmark results against other open models. That supports strong agentic capability, but the public evidence is still narrower than for the most established frontier families. | Meta | 7.9/10 AI review detailsMedium confidence · Oct 17, 2026 Coding: 7.7/10 Reasoning: 8.1/10 Reliability: 7.7/10 Instructions: 8/10 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $0.300 | $1.200 | 131k | Multimodal text+image input, Tool-calling for agents, Persistent local-agent design, Competitive benchmark evidence | Always-on local agents, Local multimodal assistants, Tool-using productivity workflows, Lightweight coding support |
| Google: Gemini 3.5 Flash Lite Listed Jul 21, 2026 Google’s model card gives detailed evaluations and limitations for Gemini 3.5 Flash-Lite, showing strong agentic and coding performance plus native multimodal support. The evidence is official and unusually complete, though it remains a fast/cost-efficient tier rather than a frontier flagship. | 7.9/10 AI review detailsLow confidence · Oct 17, 2026 Coding: 7.6/10 Reasoning: 8.1/10 Reliability: 8/10 Instructions: 8/10 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $0.300 | $2.500 | 1M | Strong agentic and coding results, Native multimodal input support, Large 1M-token context window, Well-documented limitations and evaluations | High-throughput assistant tasks, Agentic workflows, Coding and terminal tasks, Document, image, audio, and video analysis | |
| Google: Gemini 3.5 Flash Lite (batch) Listed Jul 21, 2026 This appears to be the batch routing variant of Gemini 3.5 Flash-Lite rather than a distinct model. I therefore score it the same as the base model, but with lower confidence because the batch alias itself is not separately documented in the sources I found. | 7.9/10 AI review detailsLow confidence · Oct 17, 2026 Coding: 7.6/10 Reasoning: 8.1/10 Reliability: 8/10 Instructions: 8/10 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $0.150 | $1.250 | 1M | Batch route for the same underlying model, Strong agentic and coding results, Native multimodal input support, Large 1M-token context window | High-throughput assistant tasks, Agentic workflows, Coding and terminal tasks, Document, image, audio, and video analysis | |
| OpenAI: GPT-5.6 Terra Listed Jul 9, 2026 Terra is OpenAI’s balanced GPT-5.6 tier and sits below Sol. It is a practical high-capability model, but not the family flagship. | OpenAI | 7.9/10 | $2.000 | $12.00 | 1.1M | Balanced general-purpose tier, Good coding and reasoning, Large context window, Practical everyday assistant | Everyday coding, Reasoning tasks, Document analysis, Agent workflows |
| OpenAI: GPT-5.6 Terra (batch) Listed Jul 9, 2026 Terra is OpenAI’s balanced GPT-5.6 tier and sits below Sol. It is a practical high-capability model, but not the family flagship. | OpenAI | 7.9/10 | $1.000 | $6.000 | 1.1M | Balanced general-purpose tier, Good coding and reasoning, Large context window, Practical everyday assistant | Everyday coding, Reasoning tasks, Document analysis, Agent workflows |
| Tencent: Hy3 Listed Jul 6, 2026 Tencent’s official Hy3 materials and release notes present it as a stronger business, coding, and agent model than prior Hunyuan generations. Evidence is mostly vendor-provided, so I treat the score as strong but not independently benchmark-verified. | Tencent | 7.9/10 | $0.132 vs prev +0.050 | $0.528 vs prev +0.198 | 262k | Positioned by Tencent for agentic workflows, Reported gains in coding and long-form reasoning, Official support for production use, Large context window | agentic task execution, software engineering assistance, long-document analysis, business workflow automation |
| Anthropic: Claude Sonnet 5 Listed Jun 30, 2026 Competent and still strong, but Anthropic’s newer 5.5 release is a clear improvement. Best viewed as a solid earlier generation rather than a current top pick. | Anthropic | 7.9/10 | $2.000 | $10.00 | 1M | capable Sonnet-class performance, good for scoped work, solid coding, strong professional writing | everyday assistant tasks, feature implementation, drafting and editing, structured professional work |
| Anthropic: Claude Sonnet 5 (batch) Listed Jun 30, 2026 Competent and still strong, but Anthropic’s newer 5.5 release is a clear improvement. Best viewed as a solid earlier generation rather than a current top pick. | Anthropic | 7.9/10 | $1.000 | $5.000 | 1M | capable Sonnet-class performance, good for scoped work, solid coding, strong professional writing | everyday assistant tasks, feature implementation, drafting and editing, structured professional work |
| Z.ai: GLM 5.2 Listed Jun 16, 2026 GLM 5.2 appears to be a strong long-context reasoning model with broad coding and agent-work support. Public evidence suggests solid capability, though not enough to put it clearly above the best frontier peers. | Z.ai | 7.9/10 | $0.620 | $2.420 | 1M | Very long context, Good project-level workflows, Strong reasoning profile, Low cost | Long-horizon agents, Project coding, Large document analysis, Batch reasoning |
| Qwen: Qwen3.7 Plus Listed Jun 3, 2026 Qwen’s official pages identify Qwen3.7-Plus as a flagship multimodal model with 1M context, and Qwen’s benchmark pages show solid results on coding, instruction following, and multimodal agent tasks. The evidence is decent but less definitive than a full technical report, so I rate confidence as medium. | Qwen | 7.9/10 | $0.320 | $1.280 | 1M | Strong multimodal general-purpose model, 1,000,000-token context, Good coding and agentic benchmark results in Qwen materials, Supports text, image, and video inputs | Long-context multimodal assistants, Coding and agentic workflows, Instruction-heavy enterprise applications |
| Qwen: Qwen3.7 Plus (batch) Listed Jun 3, 2026 Official Alibaba docs position Qwen3.7-Plus as a flagship multimodal model with 1M context, tools, and structured output. Evidence for broad capability is solid, but public exact-benchmark coverage is limited and there are still version-specific tradeoffs versus newer Qwen releases. | Qwen | 7.9/10 | $0.230 | $0.920 | 262k | Native multimodal input with text, images, and video; text output only., Supports function calling, built-in tools, and structured JSON output., 1M-context class model with large output budget, useful for long workflows., Official docs describe it as a flagship Qwen model with hybrid-thinking support. | Long-context multimodal analysis and extraction., Agentic workflows that need tool use and function calling., Structured data generation from documents, screenshots, or images., General-purpose reasoning and drafting where a strong proprietary model is acceptable. |
| Google: Gemini 3.5 Flash Listed May 19, 2026 Gemini 3.5 Flash is positioned as a high-efficiency model with near-Pro capability in some workflows. Flash-tier tradeoffs keep it below the best reasoning models, but it remains very strong and practical. | 7.9/10 | $1.500 | $9.000 | 1M | high-efficiency multimodal model, strong practical coding, fast parallel agent execution, good price-performance | high-throughput automation, coding assistance, multimodal apps, cost-sensitive agents | |
| Google: Gemini 3.5 Flash (batch) Listed May 19, 2026 Gemini 3.5 Flash is positioned as a high-efficiency model with near-Pro capability in some workflows. Flash-tier tradeoffs keep it below the best reasoning models, but it remains very strong and practical. | 7.9/10 | $0.750 | $4.500 | 1M | high-efficiency multimodal model, strong practical coding, fast parallel agent execution, good price-performance | high-throughput automation, coding assistance, multimodal apps, cost-sensitive agents | |
| SpaceXAI: Grok 4.3 Listed Apr 30, 2026 Grok 4.3 appears to be a strong reasoning model with good tool use and instruction following. The public evidence is mostly from xAI, so I keep confidence moderate and reliability a bit lower than the top models. | xAI | 7.9/10 | $1.250 | $2.500 | 1M | Good reasoning, Useful instruction following, Multimodal input support, Agentic workflows | General assistant tasks, Coding help, Instruction-heavy workflows, Image-grounded analysis |
| SpaceXAI: Grok 4.3 (batch) Listed Apr 30, 2026 Grok 4.3 appears to be a strong reasoning model with good tool use and instruction following. The public evidence is mostly from xAI, so I keep confidence moderate and reliability a bit lower than the top models. | xAI | 7.9/10 | $1.000 | $2.000 | 1M | Good reasoning, Useful instruction following, Multimodal input support, Agentic workflows | General assistant tasks, Coding help, Instruction-heavy workflows, Image-grounded analysis |
| Mistral: Mistral Medium 3.5 Listed Apr 30, 2026 Mistral Medium 3.5 is positioned as a frontier multimodal model aimed at agentic and coding work. The evidence is solid, but still largely vendor-led, so I rate it strong rather than undisputed frontier. | Mistral | 7.9/10 | $1.500 | $7.500 | 262k | Strong multimodal support, Agentic workflow oriented, Good coding utility, Solid structured outputs | Coding support, Agent workflows, Visual tasks, Complex assistant work |
| Mistral: Mistral Medium 3.5 (batch) Listed Apr 30, 2026 Mistral Medium 3.5 is positioned as a frontier multimodal model aimed at agentic and coding work. The evidence is solid, but still largely vendor-led, so I rate it strong rather than undisputed frontier. | Mistral | 7.9/10 | $0.750 | $3.750 | 262k | Strong multimodal support, Agentic workflow oriented, Good coding utility, Solid structured outputs | Coding support, Agent workflows, Visual tasks, Complex assistant work |
| DeepSeek: DeepSeek V4 Pro 0423 Listed Apr 24, 2026 DeepSeek V4 Pro is presented as a large-scale MoE model for advanced reasoning and coding. The model looks capable, but the evidence base is still more vendor-published than independently validated. | DeepSeek | 7.9/10 | $0.209 | $0.418 | 1M | Very large context window, Strong reasoning and coding emphasis, Cost-efficient relative to top flagships, Tool and structured-output support | Long-context reasoning, Coding tasks, Budget-conscious advanced use, Agentic workflows |
| Xiaomi: MiMo-V2.5 Listed Apr 22, 2026 Xiaomi’s official materials describe MiMo-V2.5 as a native omni-modal agent model with strong practical task focus. It appears capable and well-rounded, but the public evidence is more descriptive than deeply benchmark-heavy. | Xiaomi | 7.9/10 | $0.140 | $0.280 | 1.1M | Native omni-modal capability, Tool use and long context, Strong image/video understanding, Agent-focused design | Multimodal assistants, Document and video analysis, Agentic workflows, Long-context reasoning tasks |
| Google: Gemma 4 31B Listed Apr 2, 2026 Google’s Gemma 4 card shows the 31B instruction-tuned model performing strongly on reasoning, coding, vision, and long-context benchmarks, but it is still a smaller open model rather than a frontier proprietary flagship. That supports a strong but not frontier-wide score. | 7.9/10 | $0.200 | $0.400 | 262k | Strong open multimodal model with text and image input, Good coding and reasoning benchmarks for its size, Native function calling and system-role support, Long context and broad language support | Open-weight multimodal assistants, Coding help on self-hosted infrastructure, Document and image analysis, Agentic workflows with function calling | |
| Qwen: Qwen3.6 Plus Listed Apr 2, 2026 A meaningful step up in coding and instruction following with very large context, but not clearly above the best proprietary frontier models. It is a strong all-rounder rather than a clear leader. | Qwen | 7.9/10 | $0.325 | $1.950 | 1M | Strong all-rounder, Large context, Good coding and instruction following, Efficient inference | General productivity, Long-context tasks, Coding help, Multilingual assistance |
| MiniMax: MiniMax M2.7 Listed Mar 18, 2026 Evidence from the official model report and model card suggests a strong agentic coding model with solid reasoning and workflow skills, but most results are vendor-reported and not independently replicated here. Great fit for coding/agent tasks; less certain for broad general-purpose reliability. | MiniMax | 7.9/10 | $0.300 | $1.200 | 205k | Strong agentic software engineering and multi-step code repair, with reported gains on SWE-Pro, SWE-, Bench Multilingual, Multi-SWE-bench, and Terminal-Bench 2., Good tool- and workflow-oriented behavior for long-horizon tasks, including complex skills and agent, teams / scaffold iteration in the model report and card. Supports office-document editing and other | Autonomous coding agents, Bug fixing, log analysis, and refactoring, Long-horizon tool use and workflow automation, Office/document editing workflows |
| Qwen: Qwen3.5-122B-A10B Listed Feb 25, 2026 Official model-card evidence places Qwen3.5-122B-A10B in the strong upper tier across general, instruction, and multimodal benchmarks. It is very capable, but the public evidence does not show it clearly surpassing the very best current frontier systems. | Qwen | 7.9/10 | $0.260 | $2.080 | 262k | Strong upper-tier open model, Good multimodal and instruction performance, Large MoE efficiency, Useful for demanding workflows | Open-weight enterprise assistants, Coding and analysis, Multimodal understanding, Long-context work |
| Qwen: Qwen3.5 397B A17B Listed Feb 16, 2026 Qwen3.5 397B-A17B is presented as the open-weight flagship and appears strong across reasoning, coding, and multimodal tasks. I keep it below the strongest proprietary frontier systems because the evidence is still mostly vendor-side. | Qwen | 7.9/10 | $0.500 | $3.600 | 262k | open-weight flagship, strong reasoning and coding, multimodal support, large-scale model | open-weight deployments, research and experimentation, coding and reasoning, multimodal tasks |
| Z.ai: GLM 5 Listed Feb 11, 2026 GLM-5 looks like a strong open model for agentic engineering tasks. Public independent validation is thinner than for the leading closed systems, so I keep it below the top frontier generalists. | Z.ai | 7.9/10 | $1.000 | $3.200 | 205k | Strong agentic engineering focus, Useful for large programming tasks, Good tool-calling support, Open-weight deployment appeal | Systems design, Programming tasks, Agent workflows, Long-horizon engineering |
| StepFun: Step 3.5 Flash Listed Jan 29, 2026 Step 3.5 Flash has unusually strong primary evidence: the official paper and platform describe frontier-level agentic capability, strong coding/math/tool-use results, and efficiency-oriented architecture. It looks like a strong generalist with especially good agentic performance, though still not proven across every niche task. | StepFun | 7.9/10 | $0.100 | $0.300 | 262k | Official paper reports frontier-level agentic performance, Strong math, code, tool-use, and browsing results in the paper, Optimized for fast multi-round agent workflows, Long-context and sparse-MoE design for efficiency | Agentic coding and software engineering, Long-horizon tool-using assistants, Math/reasoning-heavy workflows, High-throughput production agents |
| MoonshotAI: Kimi K2.5 Listed Jan 27, 2026 Kimi K2.5 is described as a multimodal, agent-ready model with a strong coding emphasis. Confidence is moderate because the strongest evidence is official release material, with limited independent verification in the sources found. | MoonshotAI | 7.9/10 | $0.500 | $2.500 | 262k | multimodal agent-ready model, strong visual coding, self-directed workflow emphasis, good coding specialization | visual coding, agent swarms, multi-step programming, document/image-aware tasks |
| DeepSeek: DeepSeek V3.2 Listed Dec 1, 2025 DeepSeek V3.2 is presented as a strong general-purpose model with efficient long-context design. Most evidence is vendor-supplied, so exact placement versus leading frontier models is uncertain. | DeepSeek | 7.9/10 | $0.280 | $0.420 | 164k | efficient reasoning design, strong agentic tool use, good long-context handling, cost-effective | analysis and automation, budget deployments, structured extraction, tool-using agents |
| Qwen: Qwen3 VL 235B A22B Instruct Listed Sep 23, 2025 Qwen’s technical-report ecosystem supports this as a very strong multimodal flagship, especially for visual-language tasks. It looks frontier-adjacent for VLM work, but there is less exact public benchmark granularity for this precise instruct variant than for some other leaders. | Qwen | 7.9/10 | $0.210 | $1.900 | 262k | Very strong visual-language understanding, Good document/chart/table parsing, Large open multimodal model, Broad VQA and analysis utility | Document and chart understanding, Image/video question answering, Multimodal research workflows, Open multimodal deployment |
| DeepSeek: DeepSeek V3.1 Terminus Listed Sep 22, 2025 This looks more like a stability and polish update than a major capability leap. It should be competent and reliable, but not a top-tier frontier jump. | DeepSeek | 7.9/10 | $0.270 | $1.000 | 164k | Improved language consistency, Better agent behavior, Stable update over V3.1, Good tool use support | General chat, Agent workflows, Structured generation, Cost-sensitive reasoning |
| Qwen: Qwen3 Next 80B A3B Thinking Listed Sep 11, 2025 The exact thinking snapshot is less directly documented than the family references, so I infer capability from the broader Qwen3-Next line. That supports a strong but uncertain placement. | Qwen | 7.9/10 | $0.150 | $1.200 | 262k | Reasoning-first design, Long-context oriented family, Strong structured outputs, Efficient compared with much larger models | Hard reasoning problems, Long-context tasks, Structured analysis, Budget-conscious deployments |
| DeepSeek: DeepSeek V3.1 Listed Aug 21, 2025 DeepSeek V3.1 appears to be a strong hybrid reasoning model with improved tool use and long-context behavior. It is a top open-model contender, though reliability still trails the best proprietary systems on edge cases. | DeepSeek | 7.9/10 | $0.250 | $0.950 | 164k | Hybrid reasoning modes, Good long-context behavior, Strong tool use, Solid open-model breadth | General assistant work, Coding help, Long-context synthesis, Tool-using workflows |
| OpenAI: GPT-5 Listed Aug 7, 2025 GPT-5 was a capable frontier model, but later GPT-5.2, GPT-5.4, GPT-5.6, and GPT-6 releases supersede it. It remains useful, but it is no longer a current leader. | OpenAI | 7.9/10 | $1.250 | $10.00 | 400k | Solid frontier generalist, Good reasoning and coding, Strong instruction following, Useful baseline model | General chat and analysis, Coding help, Research summaries, Document drafting |
| OpenAI: GPT-5 (batch) Listed Aug 7, 2025 GPT-5 was a capable frontier model, but later GPT-5.2, GPT-5.4, GPT-5.6, and GPT-6 releases supersede it. It remains useful, but it is no longer a current leader. | OpenAI | 7.9/10 | $0.625 | $5.000 | 400k | Solid frontier generalist, Good reasoning and coding, Strong instruction following, Useful baseline model | General chat and analysis, Coding help, Research summaries, Document drafting |
| Anthropic: Claude Opus 4.1 Listed Aug 5, 2025 Opus 4.1 was a strong flagship, but it is clearly behind newer Anthropic releases. I keep it in the high-capability range, with the main limitation being age relative to current models. | Anthropic | 7.9/10 | $15.00 | $75.00 | 200k | Strong earlier Anthropic flagship, Good coding and reasoning, Useful agentic behavior, Still capable on long tasks | Complex coding help, Reasoning and analysis, Agent workflows, Document work |
| Anthropic: Claude Opus 4.1 (batch) Listed Aug 5, 2025 Opus 4.1 was a strong flagship, but it is clearly behind newer Anthropic releases. I keep it in the high-capability range, with the main limitation being age relative to current models. | Anthropic | 7.9/10 | $7.500 | $37.50 | 200k | Strong earlier Anthropic flagship, Good coding and reasoning, Useful agentic behavior, Still capable on long tasks | Complex coding help, Reasoning and analysis, Agent workflows, Document work |
| Qwen: Qwen3 30B A3B Instruct 2507 Listed Jul 29, 2025 The 2507 Instruct release is presented as a meaningful update with better instruction following, reasoning, coding, and tool use. It looks like a strong general-purpose model, though still below frontier top tier on the best available public evidence. | Qwen | 7.9/10 | $0.090 vs prev -0.010 | $0.300 | 262k | Strong instruction-following updates in the 2507 version, MoE model with 3.3B active params and 256K context, Improved multilingual and tool-use capabilities, Designed for broad general-purpose assistant use | General assistant tasks, Multilingual Q&A and summarization, Tool-using agent workflows, Coding assistance and reasoning tasks |
| Qwen: Qwen3 Coder 480B A35B Listed Jul 23, 2025 Qwen3-Coder is clearly a code-specialist model and should be excellent for programming tasks. It is less broad as a general assistant than frontier generalists, which keeps the overall score slightly lower. | Qwen | 7.9/10 | $0.300 | $1.000 | 262k | Excellent coding specialization, Good agentic tool use, Strong long-context reasoning for code, Efficient MoE design | Code generation, Function calling, Repo-level tasks, Long-context programming |
| Qwen: Qwen3 235B A22B Instruct 2507 Listed Jul 21, 2025 An updated general-purpose Qwen MoE model with broad capability. Public evidence is thinner than for the best-known proprietary flagships, so calibration is conservative. | Qwen | 7.9/10 | $0.090 | $0.550 | 262k | Broad general-purpose instruction tuning, Good multilingual behavior, MoE efficiency, Useful as a balanced open model | General text generation, Instruction following, Multilingual tasks, Routine coding help |
| MoonshotAI: Kimi K2 0711 Listed Jul 11, 2025 Moonshot’s official materials position Kimi K2 as a 1T-parameter MoE with 32B active parameters and strong results on coding, reasoning, and agentic benchmarks. It looks very capable, but the evidence also suggests a reflex-style model that may be less consistent than top closed reasoning systems on hard or delicate tasks. | MoonshotAI | 7.9/10 | $0.570 | $2.300 | 131k | Strong agentic/tool-use orientation, Competitive coding performance in the official report, Large context window and open weights, Good general reasoning on knowledge/math tasks | Agentic coding and software patches, Long-context analysis and retrieval-heavy tasks, Tool-using assistants and workflow automation, Open-weight deployment and fine-tuning |
| Google: Gemini 2.5 Flash Listed Jun 17, 2025 Gemini 2.5 Flash is a strong thinking model for high-volume tasks, with stable 2.5-era reasoning behavior. | 7.9/10 | $0.300 | $2.500 | 1M | Very strong price-performance, Good reasoning for high-volume use, Large context window, Multimodal and tool-capable | High-volume assistants, Reasoning at scale, Budget-conscious coding help, Multimodal extraction | |
| Google: Gemini 2.5 Flash (batch) Listed Jun 17, 2025 Gemini 2.5 Flash is a strong thinking model for high-volume tasks, with stable 2.5-era reasoning behavior. | 7.9/10 | $0.150 | $1.250 | 1M | Very strong price-performance, Good reasoning for high-volume use, Large context window, Multimodal and tool-capable | High-volume assistants, Reasoning at scale, Budget-conscious coding help, Multimodal extraction | |
| Google: Gemini 2.5 Pro Preview 06-05 Listed Jun 5, 2025 Gemini 2.5 Pro Preview is close to the stable model but less dependable because it is a preview. I keep capability high and reliability slightly lower to reflect that status. | 7.9/10 | $1.250 | $10.00 | 1M | Strong preview reasoning model, Good multimodal support, Useful for coding and science, Large context window | Testing latest Gemini capabilities, Advanced reasoning, Multimodal tasks, Long-document analysis | |
| DeepSeek: R1 0528 Listed May 28, 2025 DeepSeek R1-0528 is a strong reasoning model with clear improvements over earlier R1 releases. It is very capable, though the documentation and external validation are still less detailed than the top closed-model cards. | DeepSeek | 7.9/10 | $0.500 | $2.150 | 164k | Strong reasoning depth, Open reasoning tokens, Good tool behavior, Competitive open model | Hard reasoning, Math and logic, Agentic workflows, Research assistance |
| OpenAI: GPT-4.1 Listed Apr 14, 2025 A capable older OpenAI flagship focused on instruction following and long-context work. It predates the GPT-5/6 generations, so it is no longer near the top of the stack. | OpenAI | 7.9/10 | $2.000 | $8.000 | 1M | Strong instruction following, Good long-context reasoning, Solid software engineering, Reliable older flagship | Instruction-heavy work, Long-context tasks, Software engineering, Drafting and editing |
| OpenAI: GPT-4.1 (batch) Listed Apr 14, 2025 A capable older OpenAI flagship focused on instruction following and long-context work. It predates the GPT-5/6 generations, so it is no longer near the top of the stack. | OpenAI | 7.9/10 | $1.000 | $4.000 | 1M | Strong instruction following, Good long-context reasoning, Solid software engineering, Reliable older flagship | Instruction-heavy work, Long-context tasks, Software engineering, Drafting and editing |
| Mistral: Mistral Small 3.1 24B Listed Mar 17, 2025 Evidence shows a capable 24B multimodal instruct model with strong coding, reasoning, and tool-calling for its size. It is not frontier-tier overall, and the vendor benchmarks are strongest for standard evals rather than hard reliability or deep expert domains. | Mistral | 7.9/10 | $0.351 | $0.555 | 128k | Strong text reasoning for its size, with solid MMLU Pro and GPQA results in the official model card., Very good coding performance, including HumanEval and MBPP in the official benchmarks., Supports multimodal vision understanding and long-context use up to 128k tokens., Built for tool/function calling and low-latency deployment, which is useful for agentic workflows. | General-purpose assistant with mixed text-and-image inputs., Coding help, code generation, and lightweight debugging., Tool-using agents that need structured outputs or function calls., Long-document Q&A and summarization over large contexts. |
| OpenAI: o3 Mini High Listed Feb 12, 2025 OpenAI’s o3-mini is a strong small reasoning model; the high setting is a configuration rather than a separate trained SKU. | OpenAI | 7.9/10 | $1.100 | $4.400 | 200k | Strong STEM reasoning, Efficient small-model design, Good tool use, Reliable on focused tasks | Math and science problems, Cost-efficient reasoning, Tool-augmented workflows, Compact deployment scenarios |
| Mistral Large 2407 Listed Nov 19, 2024 Mistral Large 2407 is the official Large 2 release and is positioned as Mistral’s flagship model. Evidence supports a strong general-purpose model, though not clearly at the very top of the frontier. | Mistral | 7.9/10 | $2.000 | $6.000 | 131k | Strong general-purpose model, Good reasoning and JSON work, Useful function-calling ability, Multilingual strengths | General assistant, Coding support, Structured outputs, Enterprise workflows |
| Fireworks: Ember-1 Listed Sep 24, 2026 Evidence suggests Ember-1 is a specialized efficiency-tuned derivative of Kimi K3 that preserves much of K3’s quality while cutting reasoning tokens by about 40%. It looks strong for coding and agentic use, but Fireworks’ public evidence is mostly benchmark-and-customer-validation rather than a full independent model card. | Fireworks | 7.8/10 | $3.000 | $15.00 | 1M | Keeps Kimi K3-level quality with materially fewer reasoning tokens, which should help on cost- and-l, Strong fit for long-horizon coding and agentic workflows, especially where token efficiency matters., Supports native multimodal inputs and tool/function calling via the Kimi K3 base capabilities., Designed for structured output and configurable reasoning effort, making it practical for production | Agentic coding on large repositories where inference cost is a concern., Multimodal document understanding and tool-using workflows built on Kimi K3., Production copilots that need similar answer quality with shorter traces and lower spend. |
| Z.ai: GLM 5.3 FlashX Listed Sep 18, 2026 Evidence indicates FlashX is a higher-speed serving SKU of GLM-5.3-Flash rather than a new capability tier, so I score it close to Flash but with low confidence. The main limitation is that I found no independent benchmark set specific to FlashX, only launch and product docs for the serving variant and base family. | Z.ai | 7.8/10 | $0.370 | $1.250 | 1M | Native multimodal input for image-heavy workflows, Fast streamed responses suited to agent/tool loops, Strong fit for coding and long-horizon task execution, Serving-optimized variant with large context window | Coding agents and IDE copilots, Vision+text document or UI understanding, Tool-using workflow automation, Interactive assistant sessions with long context |
| DeepSeek: DeepSeek V4.1 Flash Listed Sep 10, 2026 Official DeepSeek materials present V4.1 Flash as a strong efficient model with good coding and agent behavior. It is a top-tier flash model, but the evidence reads more like a strong specialist than an unconstrained frontier generalist. | DeepSeek | 7.8/10 | $0.065 vs prev +0.008 | $0.144 | 1M | Efficient flash-tier model, Solid coding and agent performance, Multimodal support, Good cost-performance | Fast assistants, Budget-conscious coding help, Multimodal Q&A, Long-context summarization |
| DeepSeek: DeepSeek V4.1 Flash (batch) Listed Sep 10, 2026 Official DeepSeek materials present V4.1 Flash as a strong efficient model with good coding and agent behavior. It is a top-tier flash model, but the evidence reads more like a strong specialist than an unconstrained frontier generalist. | DeepSeek | 7.8/10 | $0.050 | $0.150 | 1M | Efficient flash-tier model, Solid coding and agent performance, Multimodal support, Good cost-performance | Fast assistants, Budget-conscious coding help, Multimodal Q&A, Long-context summarization |
| ByteDance Seed: Seed-2.0-Code Listed Aug 12, 2026 ByteDance’s Seed materials present Seed-2.0-Code as a specialized coding model for agentic development workflows. The evidence is credible for coding strength, but broader general capability is less well documented. | ByteDance Seed | 7.8/10 | $0.500 | $3.000 | 262k | Dedicated coding specialization, Agentic coding workflow focus, Good multilingual programming support, Designed for tool-based development | Frontend and multilingual coding, Coding agents, Repo edits and debugging, Developer tooling |
| inclusionAI: Ling 3.0 Flash Listed Jul 23, 2026 Official Ling sources describe Ling-3.0-flash as a 124B total / 5.1B active native hybrid reasoning model with strong results on code, agent, and long-context benchmarks. Most evidence is vendor-provided, but it is specific and substantial enough for a medium-confidence assessment. | inclusionAI | 7.8/10 | $0.021 | $0.063 | 262k | Strong coding/agent benchmark focus, Hybrid MoE architecture with efficient active parameters, Designed for long-context workflows, Default-thinking and production deployment emphasis | Coding agents, Deep-research workflows, Long-context document analysis, Multi-step tool-using assistants |
| MoonshotAI: Kimi K3 Listed Jul 16, 2026 Kimi K3 is presented as a strong open-weight multimodal reasoning model, and that is consistent with a solid high-end score. But compared with the most established proprietary frontier models, public independent evidence is thinner. | MoonshotAI | 7.8/10 | $0.304 | $11.94 | 1M | Strong multimodal reasoning positioning, Good coding and knowledge-work fit, Very large context window, Open-weight model with broad applicability | Coding and refactoring, Knowledge work and research, Long-context multimodal tasks, Agentic workflows |
| MoonshotAI: Kimi K3 (batch) Listed Jul 16, 2026 Kimi K3 is presented as a strong open-weight multimodal reasoning model, and that is consistent with a solid high-end score. But compared with the most established proprietary frontier models, public independent evidence is thinner. | MoonshotAI | 7.8/10 | $2.280 | $11.40 | 1M | Strong multimodal reasoning positioning, Good coding and knowledge-work fit, Very large context window, Open-weight model with broad applicability | Coding and refactoring, Knowledge work and research, Long-context multimodal tasks, Agentic workflows |
| OpenAI: GPT-5.6 Luna Pro Listed Jul 9, 2026 Evidence from OpenAI and OpenRouter shows this is GPT-5.6 Luna with reasoning.mode=pro, so capability should track Luna rather than a newer family. OpenAI positions Luna as the efficient GPT-5.6 model, with solid but not flagship performance. | OpenAI | 7.8/10 | $0.200 | $1.200 | 1.1M | Strong general reasoning for everyday and moderately complex tasks, Good tool-calling and structured-output support, Vision/image input support, Efficient enough for higher-volume use when quality still matters | Analytical drafting and synthesis, Code review and moderate coding assistance, Tool-using assistants with JSON/schema outputs, Multimodal Q&A over documents or images |
| OpenAI: GPT-5.6 Luna Pro (batch) Listed Jul 9, 2026 This batch alias is documented as the same underlying model as GPT-5.6 Luna Pro, so capability should match the non-batch version. I found no evidence that batch changes model quality; it mainly changes serving mode. | OpenAI | 7.8/10 | $0.100 | $0.600 | 1.1M | Same underlying capabilities as GPT-5.6 Luna Pro, Batch-friendly for high-throughput workloads, Tool calling and structured outputs, Image/text multimodal support | Offline batch analysis, Bulk document processing, Large-scale extraction or classification, Queued agent tasks with JSON outputs |
| OpenAI: GPT-5.6 Terra Pro Listed Jul 9, 2026 A solid balanced model, but evidence suggests a lower ceiling than GPT-5.6 Sol and GPT-6 Astra. The Pro wrapper likely increases effort more than it changes the underlying model tier. | OpenAI | 7.8/10 | $2.000 | $12.00 | 1.1M | balanced general-purpose model, good everyday professional work, likely stable and efficient, pro mode may help on hard prompts | routine business tasks, summarization and drafting, light-to-medium coding, general assistant use |
| OpenAI: GPT-5.6 Terra Pro (batch) Listed Jul 9, 2026 A solid balanced model, but evidence suggests a lower ceiling than GPT-5.6 Sol and GPT-6 Astra. The Pro wrapper likely increases effort more than it changes the underlying model tier. | OpenAI | 7.8/10 | $1.000 | $6.000 | 1.1M | balanced general-purpose model, good everyday professional work, likely stable and efficient, pro mode may help on hard prompts | routine business tasks, summarization and drafting, light-to-medium coding, general assistant use |
| SpaceXAI: Grok 4.5 Listed Jul 8, 2026 Grok 4.5 appears strong and close to the frontier, but the comparative record is not as deep as for the biggest flagship families. I score it as capable with moderate uncertainty. | xAI | 7.8/10 | $2.000 | $6.000 | 500k | Strong near-frontier model, Good coding and knowledge work, Useful STEM positioning, Large context window | Coding help, Knowledge work, STEM reasoning, General assistant tasks |
| MoonshotAI: Kimi K2.7 Code Listed Jun 12, 2026 Kimi’s docs describe K2.7 Code as an open, agentic coding model with improved benchmark scores and lower reasoning-token usage than K2.6. It looks strong for programming, but broad general-purpose evidence is thinner than for top frontier generalists. ([kimi.com](https://www.kimi.com/code/docs/en/kimi-code/whats-new.html?utm_source=openai)) | MoonshotAI | 7.8/10 | $0.671 | $3.350 | 262k | Coding-focused, Good long-horizon programming, Improved agent and instruction handling, High-speed variant available | Code completion and debugging, Long programming sessions, Agentic coding workflows, Open-model coding deployments |
| Qwen: Qwen3.6 Flash Listed Apr 27, 2026 Qwen3.6 Flash appears to be a fast, capable model with improved multimodal and code-agent performance. It looks useful for speed-sensitive workloads, though evidence is mainly vendor-reported. | Qwen | 7.8/10 | $0.188 | $1.125 | 1M | Fast and efficient, Strong multimodal input support, Useful for code-agent and reasoning tasks, Very large context | Speed-sensitive assistants, Multimodal tasks, Code-agent workflows, High-throughput prompting |
| Tencent: Hy3 preview Listed Apr 22, 2026 Tencent’s materials describe Hy3 preview as a capable agent-oriented MoE model with configurable reasoning levels. The evidence is promising, but benchmark claims are mostly vendor-reported and should be treated cautiously. | Tencent | 7.8/10 | $0.180 | $0.600 | 262k | Efficient MoE design, Configurable reasoning levels, Good agentic workflow focus, Strong cost-efficiency | Production agents, Reasoning workflows, Coding support, Low-cost automation |
| MoonshotAI: Kimi K2.6 Listed Apr 20, 2026 Moonshot’s materials present Kimi K2.6 as a next-generation multimodal agentic model for long-horizon coding. The evidence supports a strong specialist, though many claims remain vendor-published. | MoonshotAI | 7.8/10 | $0.650 | $3.410 | 262k | Strong long-horizon coding, Multimodal agentic design, Good orchestration focus, Useful for UI/UX generation | End-to-end coding, UI/UX generation, Multi-agent orchestration, Long-context work |
| Z.ai: GLM 5.1 Listed Apr 7, 2026 GLM-5.1 is described by Z.ai as a major jump in coding and long-horizon task handling. The model looks strong, though there is less independent validation than for the most established frontier families. | Z.ai | 7.8/10 | $0.966 | $3.036 | 205k | Strong coding and long-horizon positioning, Tool-use support, Good agent-oriented design, Competitive efficiency | Coding tasks, Long-horizon agents, Structured tool use, General reasoning |
| SpaceXAI: Grok 4.20 Listed Mar 31, 2026 xAI’s model card and developer docs support Grok 4.20 as a strong, tool-capable multimodal model. The evidence is decent but mostly vendor-authored, so exact rank versus the strongest frontier peers remains uncertain. | xAI | 7.8/10 | $1.250 | $2.500 | 2M | Strong multimodal reasoning positioning, Good tool calling, Large context window, Useful for agentic workloads | Tool-using assistants, Long-context reasoning, Multimodal analysis, Agentic workflows |
| OpenAI: GPT-5.4 Nano Listed Mar 17, 2026 Official OpenAI docs describe GPT-5.4 nano as the cheapest GPT-5.4-class model for simple high-volume work, with multimodal and coding-oriented use cases. Evidence supports strong small-model capability, but it is still a compact variant, so edge-case depth and robustness trail larger frontier models. | OpenAI | 7.8/10 | $0.200 | $1.250 | 400k | Fast, low-cost small model, Good for high-volume classification and extraction, Supports multimodal workflows, Useful for tool-heavy subagent tasks | Large-scale triage, Structured extraction, Lightweight coding assistants, Subagents in agentic pipelines |
| OpenAI: GPT-5.4 Nano (batch) Listed Mar 17, 2026 This batch entry appears to be the same GPT-5.4 nano model exposed through a batch-serving variant, so capability should match the base model. I am less certain about deployment-specific behavior, but the underlying model evidence is strong. | OpenAI | 7.8/10 | $0.100 | $0.625 | 400k | Fast, low-cost small model, Good for high-volume classification and extraction, Supports multimodal workflows, Useful for tool-heavy subagent tasks | Large-scale triage, Structured extraction, Lightweight coding assistants, Batch subagents in agentic pipelines |
| ByteDance Seed: Seed-2.0-Lite Listed Mar 10, 2026 ByteDance’s Seed2.0 materials position Lite as a production-grade general agent with upgraded multimodal, coding, and instruction-following ability. The evidence is strong for broad capability, though the public benchmark page is more product-focused than independently replicated. | ByteDance Seed | 7.8/10 | $0.250 | $2.000 | 262k | strong multimodal understanding, agent and GUI capabilities, good long-horizon instruction handling, production-oriented general-purpose behavior | multimodal agent tasks, document/image/video understanding, GUI and workflow automation, general production assistant |
| Qwen: Qwen3.5 Plus 2026-02-15 Listed Feb 16, 2026 The hosted snapshot likely tracks the Qwen3.5 Plus family well, but the exact snapshot is harder to verify than the family-level model. I’m therefore inferring broad parity with the cited family evidence. | Qwen | 7.8/10 | $0.260 | $1.560 | 1M | capable multimodal generalist, large context, solid instruction following, efficient hosted snapshot | everyday reasoning, multimodal tasks, long-document work, budget deployments |
| Anthropic: Claude Opus 4.6 Listed Feb 4, 2026 Anthropic’s system-card index shows Opus 4.6 is an earlier release, and later Opus 5/5.5 plus Fable 5/5.1 supersede it. It remains competent, but it is no longer near the top of the stack. | Anthropic | 7.8/10 AI review detailsHigh confidence · Sep 25, 2026 Coding: 8/10 Reasoning: 7.8/10 Reliability: 7.8/10 Instructions: 7.6/10 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $5.000 | $25.00 | 1M | Still a capable Opus-class model, Strong on coding and long professional tasks, Good agentic workflow support, Useful fallback for older deployments | Legacy codebase work, General professional tasks, Agent workflows, Review and analysis |
| Anthropic: Claude Opus 4.6 (batch) Listed Feb 4, 2026 Anthropic’s system-card index shows Opus 4.6 is an earlier release, and later Opus 5/5.5 plus Fable 5/5.1 supersede it. It remains competent, but it is no longer near the top of the stack. | Anthropic | 7.8/10 AI review detailsHigh confidence · Sep 25, 2026 Coding: 8/10 Reasoning: 7.8/10 Reliability: 7.8/10 Instructions: 7.6/10 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $2.500 | $12.50 | 1M | Still a capable Opus-class model, Strong on coding and long professional tasks, Good agentic workflow support, Useful fallback for older deployments | Legacy codebase work, General professional tasks, Agent workflows, Review and analysis |
| OpenAI: GPT-5.2-Codex Listed Jan 14, 2026 OpenAI explicitly tunes this model for software engineering, so coding is its main strength. | OpenAI | 7.8/10 | $1.750 | $14.00 | 400k | Clearly code-specialized., Strong for software engineering workflows., Good for long, independent coding tasks., Competitive tool-calling and structured-output support. | Software engineering, Interactive coding sessions, Long autonomous coding tasks, Code review and implementation |
| Google: Gemini 3 Flash Preview Listed Dec 17, 2025 Gemini 3 Flash Preview is a fast, high-value thinking model with strong agentic utility. As a flash-tier preview, it should be useful broadly but sits below the very best reasoning models. | 7.8/10 | $0.500 | $3.000 | 1M | Fast thinking model, Strong agentic workflows, Good value for speed, Multimodal support | High-throughput chat, Agent workflows, Coding assistance, Multimodal tasks | |
| Google: Gemini 3 Flash Preview (batch) Listed Dec 17, 2025 Gemini 3 Flash Preview is a fast, high-value thinking model with strong agentic utility. As a flash-tier preview, it should be useful broadly but sits below the very best reasoning models. | 7.8/10 | $0.250 | $1.500 | 1M | Fast thinking model, Strong agentic workflows, Good value for speed, Multimodal support | High-throughput chat, Agent workflows, Coding assistance, Multimodal tasks | |
| OpenAI: GPT-5.2 Pro Listed Dec 10, 2025 GPT-5.2 Pro was a top-end OpenAI model, but newer 5.6 and 6 families are now above it. | OpenAI | 7.8/10 | $21.00 | $168.00 | 400k | High-end model for its generation., Strong step-by-step reasoning for routine advanced work., Useful long-context capability., Broadly capable across domains. | Advanced assistant tasks, Coding and debugging, Long-context summarization, Reasoning-heavy prompts |
| OpenAI: GPT-5.2 Pro (batch) Listed Dec 10, 2025 GPT-5.2 Pro was a top-end OpenAI model, but newer 5.6 and 6 families are now above it. | OpenAI | 7.8/10 | $10.50 | $84.00 | 400k | High-end model for its generation., Strong step-by-step reasoning for routine advanced work., Useful long-context capability., Broadly capable across domains. | Advanced assistant tasks, Coding and debugging, Long-context summarization, Reasoning-heavy prompts |
| OpenAI: GPT-5.1-Codex Listed Nov 13, 2025 GPT-5.1-Codex is an older coding-specialized model, and OpenAI’s later Codex variants are stronger. It remains useful for software tasks, but it is no longer the best choice. | OpenAI | 7.8/10 | $1.250 | $10.00 | 400k | Coding-specialized and agentic, Good for interactive dev sessions, Useful for long independent execution, Cheaper than newer Codex variants | Software engineering, Repo-level editing, Agentic coding, Interactive development |
| MoonshotAI: Kimi K2 Thinking Listed Nov 6, 2025 Kimi K2 Thinking looks promising as Moonshot’s stronger open thinking model. Independent validation is still thinner than for leading closed frontier systems, so I keep the score in the upper-strong band. | MoonshotAI | 7.8/10 | $0.600 | $2.500 | 262k | Strong open thinking model, Good long-horizon reasoning, Agentic workflow fit, Competitive cost | Reasoning tasks, Planning, Agent workflows, Coding assistance |
| Z.ai: GLM 4.6 Listed Sep 30, 2025 Official materials indicate GLM-4.6 is a meaningful improvement over GLM-4.5, but the evidence is mostly vendor-provided. I therefore keep confidence and reliability slightly below the strongest frontier models. | Z.ai | 7.8/10 | $0.500 | $2.000 | 205k | solid reasoning upgrade, tool-use support, structured outputs, competitive efficiency | general reasoning, tool-based workflows, structured generation, budget-conscious use |
| Anthropic: Claude Sonnet 4.5 Listed Sep 29, 2025 Sonnet 4.5 is a capable earlier Anthropic model, but newer Sonnet and flagship releases are better supported today. It remains useful, especially when cost or compatibility matters. | Anthropic | 7.8/10 | $3.000 | $15.00 | 1M | Good coding evidence, Useful real-world agent behavior, Large context window, Solid general capability for its class | Coding assistance, Professional drafting, Agent workflows, Multimodal review |
| Anthropic: Claude Sonnet 4.5 (batch) Listed Sep 29, 2025 Sonnet 4.5 is a capable earlier Anthropic model, but newer Sonnet and flagship releases are better supported today. It remains useful, especially when cost or compatibility matters. | Anthropic | 7.8/10 | $1.500 | $7.500 | 1M | Good coding evidence, Useful real-world agent behavior, Large context window, Solid general capability for its class | Coding assistance, Professional drafting, Agent workflows, Multimodal review |
| DeepSeek: DeepSeek V3.2 Exp Listed Sep 29, 2025 Evidence supports a strong experimental general model with long-context efficiency gains and near-parity public evals versus V3.1, but it is still explicitly experimental and not clearly a step above top frontier models. | DeepSeek | 7.8/10 | $0.270 | $0.410 | 164k | strong general reasoning for an experimental model, good long-context efficiency focus via sparse attention, solid structured-output/tool-use oriented positioning, officially positioned as near-parity with V3.1 on public evals | general-purpose assistant work with long contexts, coding help where latency/throughput matters, structured extraction and tool-orchestrated workflows |
| MoonshotAI: Kimi K2 0905 Listed Sep 4, 2025 Moonshot’s materials describe this as an update that improves agentic coding and context support while staying tool-driven. That makes it a strong coding/agent model, but not a top general reasoner. | MoonshotAI | 7.8/10 | $0.600 | $2.500 | 262k | Strong tool-driven model, Useful for agentic coding, Large-context support, Competitive open-weight option | Coding assistance, Agent workflows, Long-context tasks, Tool-using automation |
| Qwen: Qwen3 Coder 30B A3B Instruct Listed Jul 31, 2025 A top open coding model with strong agentic code performance, but it is narrower than the biggest frontier reasoning systems. I score it very well for coding and more modestly overall. | Qwen | 7.8/10 | $0.070 | $0.280 | 262k | Strong open coding model, Good repository-scale understanding, Agentic tool use, Very cost efficient | Code generation, Repo understanding, Coding agents, Budget coding workloads |
| Z.ai: GLM 4.5 Listed Jul 25, 2025 Official GLM-4.5 materials support a strong benchmark profile across reasoning, coding, and agentic tasks. It looks near-frontier for its class, but the evidence is still primarily vendor-reported. | Z.ai | 7.8/10 AI review detailsMedium confidence · Oct 17, 2026 Coding: 8/10 Reasoning: 7.9/10 Reliability: 7.4/10 Instructions: 7.7/10 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $0.600 | $2.200 | 131k | Strong flagship profile, Good reasoning and coding, Agent-focused design, Solid general utility | Agentic applications, Coding and debugging, Structured analysis, General assistant tasks |
| Google: Gemini 2.5 Pro Listed Jun 17, 2025 Google still describes Gemini 2.5 Pro as state-of-the-art for complex reasoning, code, math, STEM, and large datasets, but its own newer Gemini 3.x docs indicate it sits below the latest releases. Strong, though not current-best. | 7.8/10 | $1.250 | $10.00 | 1M | Strong stable thinking model, Large context and multimodal support, Good for code, math, and STEM, Stable release rather than preview | Complex reasoning tasks, Document and codebase analysis, Multimodal understanding, Long-context workflows | |
| Google: Gemini 2.5 Pro (batch) Listed Jun 17, 2025 Google still describes Gemini 2.5 Pro as state-of-the-art for complex reasoning, code, math, STEM, and large datasets, but its own newer Gemini 3.x docs indicate it sits below the latest releases. Strong, though not current-best. | 7.8/10 | $0.625 | $5.000 | 1M | Strong stable thinking model, Large context and multimodal support, Good for code, math, and STEM, Stable release rather than preview | Complex reasoning tasks, Document and codebase analysis, Multimodal understanding, Long-context workflows | |
| OpenAI: o4 Mini Listed Apr 16, 2025 A compact reasoning model with strong visual and math performance, but it is a smaller model so breadth is more limited than flagship systems. Best when cost and speed matter. | OpenAI | 7.8/10 | $1.100 | $4.400 | 200k | Compact and fast reasoning, Good multimodal capability, Strong cost-efficiency, Useful tool-use support | Fast reasoning tasks, Budget-conscious applications, Multimodal utility, Light agent workflows |
| OpenAI: o4 Mini (batch) Listed Apr 16, 2025 A compact reasoning model with strong visual and math performance, but it is a smaller model so breadth is more limited than flagship systems. Best when cost and speed matter. | OpenAI | 7.8/10 | $0.550 | $2.200 | 200k | Compact and fast reasoning, Good multimodal capability, Strong cost-efficiency, Useful tool-use support | Fast reasoning tasks, Budget-conscious applications, Multimodal utility, Light agent workflows |
| DeepSeek: DeepSeek V3 0324 Listed Mar 24, 2025 DeepSeek V3 0324 is a strong updated chat model with a solid technical basis and broad capability. Public reliability evidence is thinner than for some closed frontier models, so I stay somewhat conservative. | DeepSeek | 7.8/10 | $0.290 | $1.140 | 164k | Strong open chat model, Good coding performance, Competitive MoE design, Very low cost | Coding help, General chat, Long-context tasks, Cost-sensitive deployment |
| DeepSeek: DeepSeek V3 Listed Dec 26, 2024 DeepSeek-V3 is a strong general chat and coding model, but it is more of a generalist than a reasoning specialist. It is a solid open-model choice, though not a frontier leader overall. | DeepSeek | 7.8/10 AI review detailsHigh confidence · Sep 21, 2026 Coding: 8/10 Reasoning: 7.8/10 Reliability: 7.5/10 Instructions: 7.6/10 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $0.320 | $0.890 | 164k | Strong general chat model, Good coding and instruction following, Large context window, Efficient open model | General chat, Coding assistance, Long-context work, Light agent workflows |
| OpenAI: GPT-4o-mini Listed Jul 18, 2024 Official OpenAI sources show GPT-4o mini as the smaller GPT-4o family model with strong text, vision, and structured-output support. It is excellent for cost-efficient production use, but it remains clearly below flagship models on hardest reasoning tasks. | OpenAI | 7.8/10 AI review detailsHigh confidence · Oct 17, 2026 Coding: 7.5/10 Reasoning: 7.8/10 Reliability: 7.8/10 Instructions: 8/10 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $0.150 | $0.600 | 128k | Very strong small model for its size., Official support for structured outputs and tool use., Good multimodal text+image capability., Well-suited to high-volume, latency-sensitive workflows. | Cheap production assistants, Extraction and classification, Tool-calling workflows, Basic multimodal tasks |
| OpenAI: GPT-4o-mini (2024-07-18) Listed Jul 18, 2024 This dated snapshot is the officially launched GPT-4o mini model; OpenAI docs show strong text, vision, and structured-output support. It is highly capable for a small model, but still not in the flagship reasoning tier. | OpenAI | 7.8/10 | $0.150 | $0.600 | 128k | Very strong small model for its size., Official support for structured outputs and tool use., Good multimodal text+image capability., Well-suited to high-volume, latency-sensitive workflows. | Cheap production assistants, Extraction and classification, Tool-calling workflows, Basic multimodal tasks |
| OpenAI: GPT-4o (2024-05-13) Listed May 13, 2024 OpenAI released GPT-4o in the API on 2024-05-13, and the system card describes it as a high-intelligence omni model with text, image, and audio-related capabilities. It is strong, but older than later GPT-4o snapshots, so I score it a bit below the 2024-11-20 release. | OpenAI | 7.8/10 | $5.000 | $15.00 | 128k | Early GPT-4o snapshot with multimodal support, Fast, flexible assistant for common tasks, Good structured-output and tool-use support, Strong baseline for image-plus-text workflows | Chat and drafting assistance, Image understanding with text output, Structured data extraction, Light to moderate coding help |
| Nex AGI: Nex-N2.5-Pro Listed Sep 8, 2026 Nex-N2.5-Pro appears optimized for agentic coding within a visual feedback loop. I found little independent validation, so this is a cautious, low-confidence estimate. | Nex AGI | 7.7/10 | $0.075 | $0.250 | 262k | agentic coding focus, visual feedback loop, good multi-file editing, very low cost | coding agents, UI/visual iteration, codebase exploration, budget automation |
| Meta: Muse Spark 1.3 Contributor Listed Sep 2, 2026 Meta presents Muse Spark 1.3 as improved for agentic and coding tasks, but the Contributor tier has limited public evidence on exact capability deltas. It appears strong and cost-efficient, though not a top frontier closed model. | Meta | 7.7/10 | $0.100 | $0.200 | 1M | Cost-efficient multimodal reasoning, Good agentic/coding experimentation target, Large context window, Designed for collaborative, tool-using workflows | Prototype agents, Low-cost experimentation, Early-stage coding workflows, Multimodal collaboration |
| Tencent: Hy4 preview Listed Aug 28, 2026 Official Tencent sources present Hy4 preview as a 770B/49B-active open-source model with 1M+ context and productivity focus. Evidence is mostly vendor-reported, so I treat the capability claims as strong but not independently verified. | Tencent | 7.7/10 | $0.834 vs prev +0.083 | $2.501 vs prev +0.250 | 1M | Strong coding-agent focus, Large 1M+ context window, Designed for productivity workflows, Open-sourced and product-integrated | Coding agents, Office/workflow automation, Long-document analysis, Scientific/research assistance |
| Qwen: Qwen3.8 27B Listed Aug 14, 2026 The official Qwen3.8-27B materials suggest a very capable open-weight model with strong general and coding performance. It looks strong, but not clearly frontier-dominant on the hardest agentic tasks. | Qwen | 7.7/10 | $0.240 | $2.200 | 1M | Strong open-weight VLM, Good coding and research utility, Flexible thinking mode, Large-context friendly | Open-weight multimodal apps, Coding assistants, Research workflows, Long-context document work |
| ByteDance Seed: Seed 2.1 Turbo Listed Aug 12, 2026 ByteDance/BytePlus materials present Seed 2.1 Turbo as a flagship agentic model with strong coding and multimodal capabilities. The public benchmark trail is thinner than for the best-documented frontier leaders, so confidence stays moderate. | ByteDance Seed | 7.7/10 | $0.500 | $2.500 | 262k | Strong multimodal/agent positioning, Useful for coding workflows, Designed for long-horizon tasks, Good visual understanding | Software delivery workflows, Multimodal agent tasks, Visual document understanding, Task orchestration |
| NVIDIA: Nemotron 3.5 Lightning (free) Listed Aug 11, 2026 NVIDIA’s current materials describe Lightning as a 30B/3B-active hybrid model built for reasoning, long-context understanding, and agentic workloads. The evidence is strong, but it is still more specialized for efficient agent work than a top-tier all-around frontier model. | NVIDIA | 7.7/10 | $0.000 | $0.000 | 1M | Long-context agent workflows, Efficient MoE hybrid architecture, Native speculative decoding, Multilingual support | Long-running autonomous agents, Sub-agent orchestration, High-throughput tool-using workloads, General reasoning with long context |
| Thinking Machines: Inkling Listed Jul 17, 2026 Thinking Machines positions Inkling as its most capable open model, with native multimodal input and strong general-purpose and agentic use cases. Public evidence is largely vendor-provided, so capability is promising but less independently calibrated than the leading closed models. | Thinking Machines | 7.7/10 | $1.000 | $4.050 | 524k | Open-weight multimodal model with text, image, and audio, Large context window and agent/tool orientation, Strong general-purpose capability for an open model, Can be self-hosted or fine-tuned | Open-weight deployments, Multimodal assistants, Tool-use systems, Fine-tuning and research |
| Thinking Machines: Inkling (free) Listed Jul 17, 2026 Thinking Machines positions Inkling as its most capable open model, with native multimodal input and strong general-purpose and agentic use cases. Public evidence is largely vendor-provided, so capability is promising but less independently calibrated than the leading closed models. | Thinking Machines | 7.7/10 | $0.000 | $0.000 | 1M | Open-weight multimodal model with text, image, and audio, Large context window and agent/tool orientation, Strong general-purpose capability for an open model, Can be self-hosted or fine-tuned | Open-weight deployments, Multimodal assistants, Tool-use systems, Fine-tuning and research |
| OpenAI: GPT-5.6 Luna Listed Jul 9, 2026 OpenAI’s docs describe Luna as the cost-sensitive GPT-5.6 tier aimed at high-volume workloads, with structured outputs and tool support. That points to a capable production model, but not the strongest raw-capability tier in the family. | OpenAI | 7.7/10 | $0.200 | $1.200 | 1.1M | Fast and cost-efficient, Strong long-context support, Good tool support, Useful for high-volume workloads | Classification and routing, High-throughput chat, Lightweight agent workflows, Cost-sensitive production tasks |
| OpenAI: GPT-5.6 Luna (batch) Listed Jul 9, 2026 OpenAI’s docs describe Luna as the cost-sensitive GPT-5.6 tier aimed at high-volume workloads, with structured outputs and tool support. That points to a capable production model, but not the strongest raw-capability tier in the family. | OpenAI | 7.7/10 | $0.100 | $0.600 | 1.1M | Fast and cost-efficient, Strong long-context support, Good tool support, Useful for high-volume workloads | Classification and routing, High-throughput chat, Lightweight agent workflows, Cost-sensitive production tasks |
| Poolside: Laguna XS 2.1 Listed Jul 2, 2026 Poolside’s release notes and technical report frame XS 2.1 as an upgraded agentic coding model with improved SWE-bench Multilingual and terminal-bench performance. It looks specialized rather than general-purpose, with the strongest evidence in coding workflows. | Poolside | 7.7/10 | $0.060 | $0.120 | 262k | Clearly optimized for agentic coding, Strong benchmark focus on SWE-bench and terminal tasks, Open-weight and deployable locally, Good long-context support for iterative work | coding agents, repo maintenance, terminal-driven automation, local developer assistants |
| MiniMax: MiniMax M3 Listed May 31, 2026 MiniMax’s official materials present M3 as a strong multimodal foundation model for long-horizon agentic work and coding. The evidence is supportive, but some benchmark claims are self-reported, so confidence remains moderate. | MiniMax | 7.7/10 | $0.300 | $1.200 | 1M | Multimodal foundation model, Long-horizon orientation, Coding-capable, Large context window | Long-context assistant work, Coding and debugging, Multimodal workflows, Agentic task execution |
| StepFun: Step 3.7 Flash Listed May 28, 2026 StepFun’s official announcement describes Step 3.7 Flash as a high-efficiency agent model optimized for multimodal understanding, tool use, and orchestration. The model looks strong for practical agent workflows, though the benchmark claims are primarily vendor-reported rather than independently validated. | StepFun | 7.7/10 | $0.200 | $1.150 | 262k | Designed for production agent workflows, Native multimodal understanding and acting, Reliable tool use and orchestration, Openly released with ecosystem compatibility | Agentic web and desktop workflows, Tool-using multimodal assistants, Coding plus search-heavy automation |
| Qwen: Qwen3.7 Max Listed May 21, 2026 Qwen3.7-Max is positioned as a flagship model for agent-centric work, coding, and productivity. The evidence is solid, but the exact benchmark depth for this snapshot is limited. | Qwen | 7.7/10 | $1.475 | $4.425 | 1M | Strong agent-centric positioning, Good coding and productivity focus, Large context window, Useful tool-use support | Coding tasks, Office/productivity workflows, Agentic automation, General reasoning |
| Qwen: Qwen3.6 27B Listed Apr 27, 2026 Qwen3.6-27B is a capable multimodal model with strong coding and utility claims. Because it is a 27B dense model, I rate it below the largest frontier systems on broad reasoning reliability. | Qwen | 7.7/10 | $0.250 | $2.000 | 262k | Dense multimodal model, Good coding utility, Image and video inputs, Cost-efficient flagship class | Multimodal assistants, Coding support, Long-context tasks, General assistant deployment |
| Z.ai: GLM 5V Turbo Listed Apr 1, 2026 GLM-5V Turbo appears well suited for multimodal agent work, especially visual tasks. I would not assume it matches the best pure text reasoners outside that niche, so I rate it as strong but specialized. | Z.ai | 7.7/10 | $1.200 | $4.000 | 203k | Strong multimodal agent design, Good visual understanding, Useful for video/image tasks, Tool-driven workflows | Vision-based coding, Video analysis, Agent tasks with images, Multimodal automation |
| OpenAI: GPT-5.4 Mini Listed Mar 17, 2026 OpenAI materials present GPT-5.4 mini as a faster, more efficient model with improved reasoning, coding, and multimodal performance over GPT-5 mini. It looks strong for a small model, but public evidence is still thinner than for frontier flagships. | OpenAI | 7.7/10 | $0.750 | $4.500 | 400k | Fast and efficient, Good multimodal support, Strong for its size class, Useful for high-throughput work | Batch reasoning tasks, Light-to-medium coding help, Multimodal extraction/classification, Cost-sensitive assistant workloads |
| OpenAI: GPT-5.4 Mini (batch) Listed Mar 17, 2026 OpenAI materials present GPT-5.4 mini as a faster, more efficient model with improved reasoning, coding, and multimodal performance over GPT-5 mini. It looks strong for a small model, but public evidence is still thinner than for frontier flagships. | OpenAI | 7.7/10 | $0.375 | $2.250 | 400k | Fast and efficient, Good multimodal support, Strong for its size class, Useful for high-throughput work | Batch reasoning tasks, Light-to-medium coding help, Multimodal extraction/classification, Cost-sensitive assistant workloads |
| MiniMax: MiniMax M2.5 Listed Feb 12, 2026 MiniMax positions M2.5 as a strong real-world productivity and coding model. The evidence supports a capable upper-tier system, but the public record is still mostly vendor-provided rather than independently comparative. | MiniMax | 7.7/10 | $0.270 | $0.950 | 205k | Strong productivity orientation, Good agent/workflow fit, Competitive coding focus, Large context window | Automation-heavy productivity tasks, Coding and software delivery, Multi-step business workflows, Long-context document work |
| MiniMax: MiniMax M2.1 Listed Dec 23, 2025 M2.1 looks like a strong coding-specialist model with credible agentic and instruction-following improvements over its predecessor. The public evidence is mostly vendor-reported and narrower than for top frontier generalists, so I keep confidence at medium. | MiniMax | 7.7/10 | $0.300 | $1.200 | 205k | Strong multilingual programming focus and improved software-engineering performance over M2., Evaluated on SWE-bench Verified across multiple coding-agent frameworks., Shows broad gains in code review, test generation, and instruction following., Built for real-world complex tasks rather than toy coding demos. | Coding copilots, Software-engineering agents, Test generation and code review, Multilingual app development |
| Anthropic: Claude Haiku 4.5 Listed Oct 15, 2025 Anthropic says Haiku 4.5 matches Sonnet 4 on several coding and agent tasks, supporting near-frontier small-model performance. | Anthropic | 7.7/10 | $1.000 | $5.000 | 200k | Very fast and efficient, Near-frontier small-model performance, Good coding for its size, Useful for high-volume workloads | Fast assistants, Cheap coding help, Moderation and triage, High-throughput agents |
| Anthropic: Claude Haiku 4.5 (batch) Listed Oct 15, 2025 Anthropic says Haiku 4.5 matches Sonnet 4 on several coding and agent tasks, supporting near-frontier small-model performance. | Anthropic | 7.7/10 | $0.500 | $2.500 | 200k | Very fast and efficient, Near-frontier small-model performance, Good coding for its size, Useful for high-volume workloads | Fast assistants, Cheap coding help, Moderation and triage, High-throughput agents |
| OpenAI: gpt-oss-20b (batch) Listed Aug 5, 2025 OpenAI describes gpt-oss-20b as a 21B-parameter open-weight MoE reasoning model with 3.6B active parameters and strong tool use. It looks capable and practical, but it is still a smaller model than frontier systems, so depth and robustness are limited versus the largest models. | OpenAI | 7.7/10 | $0.024 | $0.112 | 131k | Strong reasoning for its size, Good tool-use orientation, Open-weight and easy to deploy locally, Clear Apache 2.0 licensing and usage docs | Local or edge deployment, Tool-using assistants, Rapid prototyping and fine-tuning, Reasoning-focused workflows with moderate compute |
| Google: Gemma 3 27B Listed Mar 12, 2025 Google’s official Gemma 3 technical report and model card show Gemma-3-27B-IT as a strong multimodal instruction model with improved math, chat, instruction-following, and multilingual performance. It is highly capable for an open model, but still not a clear frontier generalist versus top proprietary systems. | 7.7/10 | $0.080 | $0.350 vs prev -0.100 | 131k | Strong multimodal instruction model, Officially improved math, chat, instruction-following, and multilingual ability, Large 27B open model with extensive technical documentation, Good fit for vision-language and assistant workflows | Multimodal assistants, Document and image understanding, General chat and writing, Moderate coding and reasoning tasks | |
| Qwen2.5 Coder 32B Instruct Listed Nov 11, 2024 Qwen’s official materials present this as a flagship code-specialist model with strong code-generation and reasoning performance. It is clearly strong for coding, but the evidence is narrower for broad general-purpose robustness. | Qwen | 7.7/10 | $0.660 | $1.000 | 33k | Strong code generation, Strong code reasoning, Specialized coding focus, Good open-model deployability | Coding copilots, Code review and refactoring, Programming education, Code-centric agents |
| Mistral Large Listed Feb 26, 2024 Mistral Large 2 is a solid older flagship-class model with credible strengths in reasoning, code, and structured outputs. It is competent, but clearly behind current frontier leaders. | Mistral | 7.7/10 | $2.000 | $6.000 | 128k | Solid general reasoning, Good JSON and chat work, Tool-use support, Established flagship lineage | General assistant work, JSON and structured tasks, Multilingual tasks, Coding support |
| OpenAI: GPT-6 Luna Pro Listed Sep 22, 2026 OpenAI describes Luna as the efficient GPT-6 family model, and Pro mode adds more reasoning effort for difficult prompts. Capability looks solid but clearly below Astra and the other frontier-heavy options, with limited public evidence for the exact Pro wrapper. | OpenAI | 7.6/10 AI review detailsMedium confidence · Sep 29, 2026 Coding: 7.4/10 Reasoning: 7.8/10 Reliability: 7.5/10 Instructions: 7.6/10 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $0.100 | $0.500 | 1.1M | Efficient general-purpose model, Good for high-volume, focused tasks, Supports tools and multimodal workflows, Pro mode may help on harder prompts | Cost-sensitive production workloads, Routine analysis and drafting, High-throughput tool-using agents, Everyday assistant tasks |
| OpenAI: GPT-6 Luna Pro (batch) Listed Sep 22, 2026 OpenAI describes Luna as the efficient GPT-6 family model, and Pro mode adds more reasoning effort for difficult prompts. Capability looks solid but clearly below Astra and the other frontier-heavy options, with limited public evidence for the exact Pro wrapper. | OpenAI | 7.6/10 AI review detailsMedium confidence · Sep 29, 2026 Coding: 7.4/10 Reasoning: 7.8/10 Reliability: 7.5/10 Instructions: 7.6/10 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $0.050 | $0.250 | 1.1M | Efficient general-purpose model, Good for high-volume, focused tasks, Supports tools and multimodal workflows, Pro mode may help on harder prompts | Cost-sensitive production workloads, Routine analysis and drafting, High-throughput tool-using agents, Everyday assistant tasks |
| PrismML: Ternary Bonsai 2 27B Listed Sep 18, 2026 Evidence supports a strong compact multimodal reasoning model derived from Qwen3.8-27B, with near-FP16 math/coding and solid instruction-following. Limits remain in vision and tool-calling relative to the strongest categories, and most evidence is vendor-published. | PrismML | 7.6/10 | $0.075 | $0.500 | 262k | Strong math/reasoning retention versus FP16 in PrismML’s own benchmark table., Native vision-language support, including images and video in the base Qwen3.8 lineage., Good coding and instruction-following on the published benchmark set., Very long native context window for long-document and repository-style tasks. | Long-context analysis and document Q&A., Coding assistance and code generation., Image/document understanding workflows., On-device or offline agent-style use where a compact multimodal model is preferred. |
| Sakana: Fugu Ultra v2 Listed Sep 11, 2026 Sakana’s materials present Fugu Ultra v2 as the higher-capability orchestration model in the family, with strong reported benchmark results. The setup is interesting and capable, but generalization beyond the reported tasks is still uncertain. | Sakana | 7.6/10 AI review detailsMedium confidence · Oct 17, 2026 Coding: 7.7/10 Reasoning: 7.9/10 Reliability: 6.9/10 Instructions: 7.3/10 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $5.000 | $30.00 | 1M | Agent-orchestration approach, Strong hard-benchmark story, Useful for multi-agent research, Large-context setup | Research into agentic orchestration, Hard benchmark tasks, Coding workflow experiments, Long-horizon planning |
| Nex AGI: Nex-N2.5-Mini Listed Sep 8, 2026 Agentic, tool-using mini model oriented around browser and computer-use workflows. | Nex AGI | 7.6/10 | $0.025 | $0.100 | 262k | Strong agentic workflow focus with computer-use, browser-use, and visual feedback loops. | — |
| Upstage: Solar Pro 4 Listed Aug 10, 2026 Upstage positions Solar Pro 4 as a cost-efficient long-context model for complex commercial workflows. The evidence supports strong practical utility, but not a claim of frontier-leading coding or general reasoning. | Upstage | 7.6/10 | $0.090 | $0.360 | 524k | Strong long-context handling, Good office/productivity workflows, Efficient for document-heavy use, Agentic workflow orientation | Document processing, Business productivity tasks, Long-horizon office automation, Cost-conscious enterprise assistants |
| DeepSeek: DeepSeek V4 Flash 0731 Listed Jul 31, 2026 DeepSeek V4 Flash 0731 has credible official evidence for strong agent and coding performance, but the exact revision’s public comparison set is narrower than for the top Western frontier models. That makes it a capable upper-mid choice with moderate uncertainty. | DeepSeek | 7.6/10 | $0.100 vs prev +0.092 | $0.300 vs prev -0.980 | 1M | Strong public-agent benchmark claims, Very long context window, Re-post-trained revision of the V4-Flash line, Low-cost high-throughput positioning | Budget coding agents, Long-context retrieval and summarization, Tool-heavy automation, General reasoning at scale |
| Anthropic: Claude Fable 5 Listed Jun 9, 2026 Anthropic’s Fable materials indicate a frontier-leaning model family for hard knowledge work, and the available evidence suggests strong agentic capability; however, the exact ceiling of Fable 5 is less cleanly isolated than newer variants, so confidence is lower. ([anthropic.com](https://www.anthropic.com/claude/fable)) | Anthropic | 7.6/10 | $10.00 | $50.00 | 1M | Built for autonomous knowledge work and coding, Handles long, multi-step tasks, Supports file/image workflows, Likely strong on structured agentic execution | Autonomous research workflows, Long coding sessions, Knowledge-work automation, File- and image-assisted tasks |
| Anthropic: Claude Fable 5 (batch) Listed Jun 9, 2026 Anthropic’s Fable materials indicate a frontier-leaning model family for hard knowledge work, and the available evidence suggests strong agentic capability; however, the exact ceiling of Fable 5 is less cleanly isolated than newer variants, so confidence is lower. ([anthropic.com](https://www.anthropic.com/claude/fable)) | Anthropic | 7.6/10 | $5.000 | $25.00 | 1M | Built for autonomous knowledge work and coding, Handles long, multi-step tasks, Supports file/image workflows, Likely strong on structured agentic execution | Autonomous research workflows, Long coding sessions, Knowledge-work automation, File- and image-assisted tasks |
| SpaceXAI: Grok Build 0.1 Listed May 20, 2026 xAI’s official release and docs describe Grok Build 0.1 as a coding-specialist model for agentic software workflows. That supports high coding scores, while general reasoning and reliability remain a bit below frontier generalists. | xAI | 7.6/10 | $1.000 | $2.000 | 256k | Explicitly trained for agentic coding tasks, Official docs emphasize web dev, debugging, and MCP support, Fast coding-oriented model with tool-calling and structured outputs, Model is clearly documented by xAI with exact API name | Agentic software engineering, Debugging and code repair, Rapid code generation, Tool-using coding assistants |
| Qwen: Qwen3.5 Plus 2026-04-20 Listed Apr 27, 2026 This exact dated slug is documented on OpenRouter, while official Qwen sources cover the broader Qwen3.5 family and its multimodal/thinking-oriented design. I can assess it as strong, but the exact dated variant lacks direct primary benchmark evidence. | Qwen | 7.6/10 | $0.300 | $1.800 | 1M | Large 1M-token context on OpenRouter listing, Family docs indicate strong multimodal and thinking-enabled Qwen3.5 capability, Useful for coding and agentic workflows in the Qwen3.5 line, Broadly capable long-context assistant | Long-context retrieval and summarization, Coding assistance, Multimodal Q&A, Agent workflows with reasoning |
| Qwen: Qwen3.6 35B A3B Listed Apr 27, 2026 Official Hugging Face and deployment docs identify the exact 35B/3B multimodal MoE checkpoint and its long-context, tool-calling setup. Evidence is strong for broad competence, though it is still a compact open model with expected gaps versus frontier giants. | Qwen | 7.6/10 | $0.150 | $1.000 | 262k | Open-weight multimodal MoE model, 35B total / ~3B active parameters with 262K context, Native tool calling and reasoning support in public docs, HF model card and config expose the exact checkpoint | General multimodal assistant, Agentic coding and repo-level reasoning, Long-context extraction and analysis, Tool-using structured-output workflows |
| Z.ai: GLM 5 Turbo Listed Mar 15, 2026 Official Z.ai/AutoClaw docs describe GLM-5-Turbo as optimized for continuous agent workflows and tool collaboration, but public evidence is lighter than for top frontier flagships. Strong for practical agent work; less clear on broad benchmark dominance. | Z.ai | 7.6/10 | $1.200 | $4.000 | 203k | agent/workflow orchestration, tool use and structured outputs, long-context task execution, fast iterative handling of multi-step tasks | workflow automation, multi-tool research and processing, business process execution, agentic task decomposition |
| NVIDIA: Nemotron 3 Super Listed Mar 11, 2026 NVIDIA’s technical report and docs support Nemotron 3 Super as a strong open hybrid model tuned for agentic reasoning. It looks excellent for workflow agents, though not uniformly frontier-leading across every benchmark dimension. | NVIDIA | 7.6/10 | $0.080 | $0.450 | 262k | Efficient open hybrid architecture, Strong multi-agent orientation, Good compute-to-quality tradeoff, Suitable for agent scaffolding | Open-weight agent systems, Reasoning-heavy automation, Research deployments, Cost-conscious inference |
| NVIDIA: Nemotron 3 Super (free) Listed Mar 11, 2026 NVIDIA’s technical report and docs support Nemotron 3 Super as a strong open hybrid model tuned for agentic reasoning. It looks excellent for workflow agents, though not uniformly frontier-leading across every benchmark dimension. | NVIDIA | 7.6/10 | $0.000 | $0.000 | 262k | Efficient open hybrid architecture, Strong multi-agent orientation, Good compute-to-quality tradeoff, Suitable for agent scaffolding | Open-weight agent systems, Reasoning-heavy automation, Research deployments, Cost-conscious inference |
| Z.ai: GLM 4.7 Listed Dec 22, 2025 Z.ai’s official materials and README indicate a serious coding-and-agents model with thoughtful reasoning controls. It looks strong and competitive, but current evidence is mostly vendor-authored, so exact frontier rank remains uncertain. | Z.ai | 7.6/10 | $0.540 | $1.980 | 205k | Strong coding focus with interleaved/preserved thinking, Good agentic workflow controls, Open-model deployment flexibility, Relevant benchmark reporting for coding and agent tasks | Coding agents and software engineering, Long-horizon multi-turn tasks, Self-hosted or open deployment environments, Research and evaluation of reasoning control methods |
| NVIDIA: Nemotron 3 Nano 30B A3B Listed Dec 14, 2025 NVIDIA’s primary sources show Nemotron 3 Nano as an efficient, well-documented model with strong reasoning and agentic design. I read it as a specialist efficiency play rather than a universal top-tier frontier model. | NVIDIA | 7.6/10 | $0.060 | $0.240 | 262k | Very compute-efficient, Strong agentic specialization, Long-context support, Good open-model infrastructure fit | Efficient specialized agents, Developer tooling, Long-context assistants, Cost-sensitive deployments |
| OpenAI: GPT-5.2 Listed Dec 10, 2025 GPT-5.2 was a strong OpenAI flagship in its generation, but it now sits below the newer 5.6 and 6 families. | OpenAI | 7.6/10 | $1.750 | $14.00 | 400k | Strong broad model for its generation., Good agentic and long-context performance., Competent across common professional tasks., Solid baseline for analysis and drafting. | General professional assistant work, Moderate coding tasks, Long-context summarization, Routine reasoning tasks |
| OpenAI: GPT-5.2 (batch) Listed Dec 10, 2025 GPT-5.2 was a strong OpenAI flagship in its generation, but it now sits below the newer 5.6 and 6 families. | OpenAI | 7.6/10 | $0.875 | $7.000 | 400k | Strong broad model for its generation., Good agentic and long-context performance., Competent across common professional tasks., Solid baseline for analysis and drafting. | General professional assistant work, Moderate coding tasks, Long-context summarization, Routine reasoning tasks |
| Mistral: Devstral 2 2512 Listed Dec 9, 2025 This is a strong coding-specialist model, with official docs emphasizing tool use, codebase exploration, and multi-file editing. It should score high on coding, but only moderately on general-purpose reasoning and instruction-following compared with broader frontier models. | Mistral | 7.6/10 | $0.400 | $2.000 | 262k | Explicitly specialized for software engineering and code agents., Designed to use tools to explore codebases and edit multiple files., 256k context and support for function calling, structured outputs, batching., Official docs call it frontier code-agents model; deprecated in favor of newer Mistral model. | SWE agents that inspect and modify repositories., Multi-file code refactors and bug fixing., Tool-using codebase exploration., Automated coding workflows with long context. |
| Nous: Hermes 4 405B Listed Aug 26, 2025 Hermes 4 405B shows strong open-model capability, with good results on math, reasoning, and instruction-following. The evidence is solid but older and more self-reported than the closed-model competitors, so confidence is lower. | Nous | 7.6/10 | $1.000 | $3.000 | 131k | Strong open-weight reasoning model with hybrid thinking mode, Good instruction-following and structured-output behavior, Solid math/logic/code results in the technical report, Tool-use and function-calling support are explicit strengths | Open-weight local or self-hosted deployments, Reasoning-heavy tasks with controllable deliberation, Structured output and function calling workflows, Research or eval setups that need inspectable model behavior |
| OpenAI: gpt-oss-120b Listed Aug 5, 2025 OpenAI positions gpt-oss-120b as a strong open-weight model, but it is text-only and less broad than multimodal proprietary systems. I score it as capable but not at the very top overall. | OpenAI | 7.6/10 | $0.150 | $0.600 | 131k | Strong open-weight reasoning, Useful for production deployment, Good agentic design, Cost-efficient | Open-weight deployments, Reasoning tasks, Agentic assistants, General text workflows |
| Mistral: Mistral Medium 3 Listed May 7, 2025 Mistral’s docs describe Medium 3 as a multimodal flagship model with 128K context and built-in structured outputs, function calling, agents, and batch support. Strong general-purpose enterprise model, though newer Medium 3.5 is recommended for new integrations. | Mistral | 7.6/10 AI review detailsHigh confidence · Oct 17, 2026 Coding: 7.4/10 Reasoning: 7.8/10 Reliability: 7.6/10 Instructions: 7.7/10 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $0.400 | $2.000 | 131k | Frontier-class multimodal general model, Structured outputs and function calling, Agents/batch/tooling support, Large 128K context window | Enterprise chat assistants, Multimodal document workflows, Tool-using agents, Structured extraction and orchestration |
| OpenAI: o4 Mini High Listed Apr 16, 2025 o4-mini-high should be stronger than the base compact model because it is the high-effort variant. It is still a small model, so I keep it below the frontier generalists. | OpenAI | 7.6/10 | $1.100 | $4.400 | 200k | Very capable compact reasoner, Fast and cost-efficient, Good structured outputs, Reasoning-effort control | Lightweight reasoning, STEM work, Coding on a budget, High-throughput agents |
| DeepSeek: R1 Listed Jan 20, 2025 DeepSeek’s paper says R1 matches OpenAI o1-1217 on reasoning tasks and is designed via large-scale RL. The same evidence and later analyses also point to readability, language-mixing, and rumination drawbacks, so it is strong at reasoning but less balanced overall. | DeepSeek | 7.6/10 | $0.700 | $2.500 | 64k | Excellent reasoning orientation, Open reasoning model with strong benchmark evidence, Useful for math and multi-step deliberation, Can be distilled into smaller variants for deployment | Hard reasoning and math tasks, Research prototypes needing visible chain-of-thought style behavior, Open-model experimentation and distillation, Analysis tasks where deliberation matters more than brevity |
| Microsoft: Phi 4 Listed Jan 10, 2025 Microsoft’s model card describes Phi-4 as a 14B dense model aimed at reasoning, logic, and constrained environments, with supervised fine-tuning and DPO for alignment. The official card plus the technical report give solid evidence of strong small-model general capability, though it is not a frontier-scale model. | Microsoft | 7.6/10 | $0.070 | $0.140 | 16k | Strong compact reasoning model, Well-documented 14B dense architecture, Good instruction adherence for its size, Suitable for memory/latency-constrained deployment | Reasoning assistants, On-device or low-memory deployments, General chat and tutoring, Moderate coding and analysis tasks |
| NVIDIA: Nemotron 3.5 Super VL (free) Listed Oct 15, 2026 Official NVIDIA docs and the launch report support strong multimodal reasoning and agentic use, but public evidence is still mostly vendor-provided and focused more on broad benchmarks than coding-specific evaluations. I’m moderately confident it is a strong general-purpose multimodal model, with more limited evidence for top-tier coding. | NVIDIA | 7.5/10 | $0.000 | $0.000 | 1M | Strong multimodal reasoning across text, images, documents, GUI screenshots, and video., Large MoE design with 12B active parameters and long-context support up to 1M tokens., Built for tool use, agentic workflows, and structured outputs., Official NVIDIA materials indicate strong benchmark performance vs. prior large open models, but not | Document and screenshot analysis for agent workflows., Long-context multimodal assistants that need tool use and structured outputs., UI/GUI understanding and multimodal extraction tasks., General reasoning-heavy chat where multimodal inputs matter. |
| Mistral: Mistral Large 4 Listed Oct 6, 2026 Mistral’s docs and launch post show a serious frontier-class open-weight model, and Vals gives it respectable benchmark results in several enterprise-style tasks. But public independent evidence is thinner than for the Anthropic models, so exact frontier rank is less certain. | Mistral | 7.5/10 | $0.680 | $2.090 | 1M | Open-weight multimodal model, Large 1M-token context window, Competitive performance on some enterprise benchmarks, Strong tool/function-calling feature set | Private/on-prem deployments, Long-context document and multimodal workflows, Enterprise agent prototypes, Cost-sensitive production systems |
| Cohere: Command A+ Listed Sep 22, 2026 Evidence from Cohere’s docs and model card shows strong multimodal, agentic, and multilingual capabilities, but public evidence is mostly vendor-supplied and not enough to call it frontier overall. It looks strongest for enterprise workflows; coding depth and independent reliability evidence are more limited. | Cohere | 7.5/10 | $0.300 | $1.500 | 192k | Multimodal input with text and image support., Strong agentic/tool-calling orientation and structured outputs., Good multilingual coverage across 48 languages., Designed for efficient enterprise deployment at scale. | Complex enterprise agents and workflow automation., Multimodal document/chart understanding and extraction., Tool-using assistants with structured JSON outputs., Multilingual customer support and translation-adjacent tasks. |
| DeepSeek: DeepSeek V4 Pro 0813 Listed Aug 12, 2026 DeepSeek’s own docs indicate a capable MoE model with thinking and tool-use support, but most evidence is vendor-authored and the public comparative picture is mixed. It appears strong, though not clearly top-tier against the very best frontier systems. | DeepSeek | 7.5/10 | $0.660 | $1.980 | 1M | Strong cost-efficient large-model capability, Supports thinking/tool-use style workflows, Likely competitive on routine-to-advanced coding, Long-context, high-throughput use cases | Budget-sensitive coding assistants, Agent workflows with tool use, General reasoning and summarization, High-volume application backends |
| Thinking Machines: Inkling Small Listed Jul 30, 2026 Inkling-Small is presented as a smaller open-weight sibling with comparable performance on many benchmarks and lower cost/latency. It is strong for its size, but the smaller model likely trades away some robustness versus Inkling. | Thinking Machines | 7.5/10 | $0.450 | $1.200 | 524k | Efficient open-weight multimodal model, Strong performance per cost/latency, Same broad recipe as Inkling, Good for practical tool systems | Lower-cost multimodal assistants, Tool-using agents, Prototype deployments, Open-weight customization |
| Thinking Machines: Inkling Small (free) Listed Jul 30, 2026 Inkling-Small is presented as a smaller open-weight sibling with comparable performance on many benchmarks and lower cost/latency. It is strong for its size, but the smaller model likely trades away some robustness versus Inkling. | Thinking Machines | 7.5/10 | $0.000 | $0.000 | 1M | Efficient open-weight multimodal model, Strong performance per cost/latency, Same broad recipe as Inkling, Good for practical tool systems | Lower-cost multimodal assistants, Tool-using agents, Prototype deployments, Open-weight customization |
| Poolside: Laguna XS 2.1 (free) Listed Jul 2, 2026 Official Poolside docs describe Laguna XS 2.1 as a 33B-A3B agentic coding model with 256K context, and Poolside reports improved SWE-bench Multilingual and terminal-task performance over XS.2. Evidence is strongest for coding agents; broader general-purpose reasoning is less directly established. | Poolside | 7.5/10 | $0.000 | $0.000 | 262k | Agentic coding focus, 33B total / 3B active MoE efficiency, 256K context for long-horizon repo work, Strong terminal-style task orientation | Repo-level code edits and bugfixing, Terminal and tool-driven coding agents, Long-context code review and iteration |
| Anthropic: Claude Opus 4.8 Listed May 27, 2026 Opus 4.8 is still capable, but current Anthropic materials place later Fable/Opus releases above it. I treat it as solid and no longer state-of-the-art. | Anthropic | 7.5/10 | $5.000 | $25.00 | 1M | Strong older Opus release, Capable reasoning, Useful for analysis and code, Large context window | General assistant work, Coding help, Analysis tasks, Document review |
| Anthropic: Claude Opus 4.8 (batch) Listed May 27, 2026 Opus 4.8 is still capable, but current Anthropic materials place later Fable/Opus releases above it. I treat it as solid and no longer state-of-the-art. | Anthropic | 7.5/10 | $2.500 | $12.50 | 1M | Strong older Opus release, Capable reasoning, Useful for analysis and code, Large context window | General assistant work, Coding help, Analysis tasks, Document review |
| DeepSeek: DeepSeek V4 Flash 0423 Listed Apr 24, 2026 V4 Flash appears to be the efficiency-optimized member of DeepSeek’s V4 family. I score it below V4 Pro because the public materials emphasize speed and efficiency more than peak capability. | DeepSeek | 7.5/10 | $0.091 | $0.182 | 1M | Fast inference, Strong efficiency, Useful agentic design, Large context | High-throughput tasks, Cheap coding assistance, Batch workflows, Agent pipelines |
| SpaceXAI: Grok 4.20 Multi-Agent Listed Mar 31, 2026 xAI frames Grok 4.20 Multi-Agent as a collaborative agent-oriented model. The wrapper likely helps on research workflows, but the exact uplift over the base model is uncertain. | xAI | 7.5/10 | $1.250 | $2.500 | 2M | Multi-agent workflow support, Large context window, Good research-oriented orchestration, Reasoning-focused design | Deep research workflows, Collaborative agent tasks, Tool orchestration, Long-context synthesis |
| Qwen: Qwen3 VL 30B A3B Thinking Listed Oct 6, 2025 The exact model card supports a substantial multimodal reasoning model, and the thinking variant should be the strongest of the Qwen3-VL models listed here. I still leave room for gaps because public evidence is mostly repository-level rather than independently replicated benchmarks on this exact checkpoint. | Qwen | 7.5/10 | $0.290 | $1.000 | 262k | Large multimodal model with explicit thinking mode, Strong fit for complex vision-language reasoning, Supports images and video plus agent-style workflows, More capable scale than the 8B variant | Hard visual reasoning tasks, Video and document analysis, Multimodal agent workflows, STEM-style question answering with images |
| OpenAI: GPT-5 Pro Listed Oct 6, 2025 GPT-5 Pro is now behind the newer GPT-5.6 and GPT-6 families, so it should be treated as a strong but older flagship. | OpenAI | 7.5/10 | $15.00 | $120.00 | 400k | Broadly capable general model., Good instruction-following behavior for its era., Solid coding and reasoning baseline., Useful for routine professional tasks. | General Q&A, Drafting and editing, Moderate coding tasks, Routine analysis |
| OpenAI: GPT-5 Pro (batch) Listed Oct 6, 2025 GPT-5 Pro is now behind the newer GPT-5.6 and GPT-6 families, so it should be treated as a strong but older flagship. | OpenAI | 7.5/10 | $7.500 | $60.00 | 400k | Broadly capable general model., Good instruction-following behavior for its era., Solid coding and reasoning baseline., Useful for routine professional tasks. | General Q&A, Drafting and editing, Moderate coding tasks, Routine analysis |
| Meta: Llama 4 Maverick Listed Apr 5, 2025 Meta positions Maverick as Llama 4’s most powerful open multimodal model with strong benchmark results. | Meta | 7.5/10 | $0.188 | $0.652 | 1M | Strong open multimodal capability, Good multilingual breadth, Competitive coding for an open model, Large context window | Open-model deployments, Multimodal assistants, Multilingual generation, General-purpose local/hosted use |
| Cohere: Command A Listed Mar 13, 2025 Cohere’s docs and technical report present Command A as a top-end enterprise model with strong tool use, multilinguality, and solid reasoning/code performance. | Cohere | 7.5/10 | $2.500 | $10.00 | 256k | Cohere positions it as the most performant Command model for enterprise tasks., Strong tool use, RAG, agentic workflows, and multilingual support., Large 256K context window and structured-output support., Technical report indicates strong SQL, reasoning, and code-related performance. | Enterprise agents and tool use., Retrieval-augmented generation pipelines., Multilingual business workflows., Code/SQL assistance and structured text generation. |
| Mistral: Mistral Small 3 Listed Jan 30, 2025 Official Mistral docs describe this as a 24B, low-latency instruction model with structured outputs and function calling; current public evidence supports strong small-model capability, but it is retired and less current than Mistral Small 4. Best for efficient general tasks, not frontier reasoning. | Mistral | 7.5/10 | $0.050 | $0.080 | 33k | Strong general-purpose 24B instruct model with official support for structured outputs and function/, Good fit for latency-sensitive assistant workflows, Solid small-model coding and extraction performance for its class, Apache 2.0 open-weight deployment flexibility | Structured JSON extraction, General chat assistants, Lightweight coding help, RAG-backed enterprise workflows |
| Qwen: Qwen2.5 7B Instruct Listed Oct 16, 2024 Qwen’s official materials and technical reports describe broad improvements over Qwen2, especially in coding, math, instruction following, and structured outputs. The exact 7B instruct model is a strong small generalist, though still below frontier large models on hardest reasoning tasks. | Qwen | 7.5/10 | $0.100 | $0.200 | 33k | Strong instruction following and structured-output behavior., Noted improvements in coding and mathematics over Qwen2., Solid multilingual support and long-context capability., Good general-purpose small model for tool-using applications. | Structured JSON / tool workflows, Coding and math assistance, Multilingual assistant, Long-context summarization |
| DeepSeek: DeepSeek V4.1 Flash 1017 Listed Oct 17, 2026 DeepSeek’s October 17 snapshot is described as a Pareto improvement over the previous Flash version, with stronger benchmark performance and faster completion. That suggests a solid upper-mid/frontier-adjacent model, but the public evidence remains mostly vendor-run. | DeepSeek | 7.4/10 | $0.150 | $0.600 | 1M | Improved snapshot with better coding and task completion, Efficient price-performance profile, Supports multimodal and tool-enabled workflows, Good fit for high-volume production use | Fast coding assistants, Agentic project development, Cost-sensitive multimodal workflows, High-throughput summarization and support |
| XingChen: Xing4.0 29B A4B (free) Listed Oct 16, 2026 Evidence suggests a capable MoE model with especially strong coding-agent performance and long-context/agent features. Confidence is limited by mostly vendor-reported benchmarks and little independent third-party validation. | XingChen | 7.4/10 | $0.000 | $0.000 | 262k | Strong coding-agent benchmark showing SWE-bench Verified 75.0 and Terminal-Bench 2.1 57.5 in the org, Long-context support: 256K native, extendable to 512K, Tool-calling and reasoning parsers are explicitly supported in serving configs, OpenAI-compatible deployment across vLLM and SGLang | Software engineering and code-editing assistants, Agentic workflows with tool calling, Long-context document analysis, Chinese/English general-purpose chat where self-reported benchmark strength matters |
| Reflection: Beam 1.0 501B A23B Listed Oct 16, 2026 Reflection’s materials present Beam as a strong open-weight reasoning model, especially for coding and agents. The evidence is promising but thinner than for closed frontier leaders, so broader robustness is less certain. | Reflection | 7.4/10 | $1.000 | $3.000 | 1M | Strong open-weight coding orientation, Good agentic and long-horizon workflow focus, Large sparse MoE design with efficiency emphasis, Useful as a base for fine-tuning and deployment control | Open-weight coding assistants, Agentic workflows on self-hosted infra, Fine-tuning or model adaptation, Research on efficient reasoning models |
| StepFun: Step 5 Preview Listed Oct 8, 2026 Step 5 Preview is capable and unusually strong for an open-weight release, but current evidence is mostly vendor-run and Vals places it well behind Opus 5.5 and Gemini 4 Argon overall. That makes it a solid upper-mid frontier option, not a top general-purpose leader. | StepFun | 7.4/10 | $1.000 | $2.700 | 1M | Strong open-weight agentic model with 1M context., Competitive results on several coding and finance-style benchmarks., Useful for long-horizon workflows and experimentation., Cost-effective relative to the top proprietary frontier models. | Budget-conscious agent prototyping., Long-context document analysis., Coding and workflow automation where openness matters., Finance-oriented tasks with tolerable error risk. |
| Z.ai: GLM 5.3 Prime Listed Sep 23, 2026 Evidence suggests GLM-5.3-Prime is a faster serving variant of GLM-5.3 rather than a distinct capability jump, so it may be quite strong but not clearly top frontier. Independent validation is thinner than for the leading proprietary flagships, so confidence is low. | Z.ai | 7.4/10 | $2.800 | $8.800 | 1M | High throughput variant, Large context window, Likely strong coding/agent performance for the class, Potentially cost-efficient for volume | High-throughput production workloads, Coding assistance at scale, Long-context batch tasks, Agent pipelines where speed matters |
| Anthropic: Claude Opus 4.7 Listed Apr 16, 2026 Opus 4.7 is a strong intermediate release, but later Anthropic models are better supported and likely stronger. I treat it as capable, not current-leading. | Anthropic | 7.4/10 | $5.000 | $25.00 | 1M | Capable intermediate Opus release, Good long-running agent foundation, Useful coding ability, Large context window | Long-running agents, Coding help, Analysis and review, Document work |
| Anthropic: Claude Opus 4.7 (batch) Listed Apr 16, 2026 Opus 4.7 is a strong intermediate release, but later Anthropic models are better supported and likely stronger. I treat it as capable, not current-leading. | Anthropic | 7.4/10 | $2.500 | $12.50 | 1M | Capable intermediate Opus release, Good long-running agent foundation, Useful coding ability, Large context window | Long-running agents, Coding help, Analysis and review, Document work |
| Qwen: Qwen3.5-35B-A3B Listed Feb 25, 2026 Public Qwen materials and Hugging Face listings show Qwen3.5-35B-A3B as part of the Qwen3.5 family, with multimodal and agent-oriented positioning. The public evidence supports solid general capability, but the exact A3B variant is less thoroughly documented than the flagship Qwen releases. | Qwen | 7.4/10 AI review detailsMedium confidence · Oct 17, 2026 Coding: 7.3/10 Reasoning: 7.6/10 Reliability: 7.2/10 Instructions: 7.4/10 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $0.150 | $1.000 | 262k | Large Qwen 3.5 family model, Native multimodal lineage, Strong agent-oriented positioning, Public ecosystem support on Hugging Face | General assistant use, Multimodal understanding, Coding assistance, Agent-style workflows |
| Z.ai: GLM 4.7 Flash Listed Jan 19, 2026 Public evidence is thinner than for the biggest frontier models, but Z.ai presents GLM-4.7-Flash as a 30B-class model optimized for agentic coding and planning. That supports a strong-but-not-frontier rating, with uncertainty because the evidence is mainly vendor-facing. | Z.ai | 7.4/10 | $0.060 | $0.400 | 200k | Optimized for agentic coding, Long-context support, Tool-use oriented, Efficiency-focused | Coding assistants, Agent workflows, Long-horizon task planning, Structured tool-calling tasks |
| OpenAI: GPT-5.2 Chat Listed Dec 10, 2025 OpenAI documents GPT-5.2 Chat as the ChatGPT Instant snapshot: a fast, capable workhorse with improved clarity, structure, and multimodal handling. It is competent for routine general work, but the public evidence and deprecation status argue against rating it as frontier-level. | OpenAI | 7.4/10 | $1.750 | $14.00 | 128k | Fast everyday chat model, Strong instruction-following for a lightweight model, Supports image input and structured outputs, Useful for writing, walkthroughs, and info-seeking | Low-latency chat assistants, Customer-support triage, Summarization and drafting, General-purpose multimodal Q&A |
| OpenAI: GPT-5.1-Codex-Mini Listed Nov 13, 2025 OpenAI explicitly describes this as a smaller, less-capable version of GPT-5.1-Codex. It should be strong on routine coding, but weaker than the full Codex model on harder, longer-horizon work. | OpenAI | 7.4/10 | $0.250 | $2.000 | 400k | Fast and cheaper coding helper, Good routine code tasks, Useful for interactive development, Long context for its class | Simple coding tasks, Prototype generation, Small refactors, Developer support |
| Qwen: Qwen3 VL 30B A3B Instruct Listed Oct 6, 2025 This appears to be a strong general multimodal instruct model, likely a bit better balanced than the thinking variant for everyday use. The evidence is solid but still mostly comes from the model repository and family-level technical materials, so I avoid overstating it. | Qwen | 7.4/10 | $0.150 | $0.600 | 262k | Large multimodal model for general instruction following, Strong visual understanding across images and video, Better general-use orientation than a pure reasoning variant, More capable scale than the 8B variant | Multimodal chat assistants, Screenshot and document understanding, Image-grounded instruction following, General visual QA and extraction |
| Z.ai: GLM 4.5 Air Listed Jul 25, 2025 GLM-4.5-Air keeps the same agentic design but with lower capacity than GLM-4.5, and the official benchmarks show a clear step down. It is still strong, but the evidence suggests it trails the full model on harder tasks. | Z.ai | 7.4/10 AI review detailsHigh confidence · Oct 17, 2026 Coding: 7.2/10 Reasoning: 7.7/10 Reliability: 6.8/10 Instructions: 7.4/10 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $0.130 | $0.850 | 131k | Same hybrid reasoning design as GLM-4.5, Lower-cost smaller variant for agentic workflows, Official benchmark results cover reasoning, coding, and agent use, Native function calling support | Cost-sensitive agent workflows, Reasoning with moderate complexity, Coding help when latency matters, General assistant deployment at smaller scale |
| Qwen: Qwen3 14B Listed Apr 28, 2025 The Qwen3-14B is a capable mid-size dense model, but the evidence points to competent rather than elite capability. It should be good for everyday assistants and moderate coding/reasoning, while trailing larger Qwen3 MoE and flagship models on harder tasks. | Qwen | 7.4/10 | $0.120 | $0.240 | 41k | Dense 14.8B model with hybrid thinking/non-thinking mode, Official Qwen blog says Qwen3-14B dense base matches Qwen2.5-14B-Base, 128K context in the Qwen3 family and strong multilingual support claims, Suitable for lightweight deployment relative to larger Qwen3 variants | General-purpose chat and assistant tasks, Mid-sized coding and reasoning workloads, Multilingual instruction following and translation, Applications needing a smaller local model with hybrid reasoning |
| OpenAI: o3 Listed Apr 16, 2025 o3 was a strong model, but OpenAI now lists it as deprecated for removal, so it should not be treated as current frontier leadership. It remains competent, yet its older snapshot status puts it below the newer 2026 flagship models. | OpenAI | 7.4/10 | $2.000 | $8.000 | 200k | Still a strong general reasoning model, Good technical writing and instruction following, Known behavior and stable snapshot availability, Useful where legacy reproducibility matters | Routine reasoning and analysis, Technical writing, Stable regression baselines, Non-frontier coding tasks |
| OpenAI: o3 (batch) Listed Apr 16, 2025 o3 was a strong model, but OpenAI now lists it as deprecated for removal, so it should not be treated as current frontier leadership. It remains competent, yet its older snapshot status puts it below the newer 2026 flagship models. | OpenAI | 7.4/10 | $1.000 | $4.000 | 200k | Still a strong general reasoning model, Good technical writing and instruction following, Known behavior and stable snapshot availability, Useful where legacy reproducibility matters | Routine reasoning and analysis, Technical writing, Stable regression baselines, Non-frontier coding tasks |
| Amazon: Nova Pro 1.0 Listed Dec 5, 2024 AWS describes Nova Pro as a highly capable multimodal model aimed at complex tasks, with competitive accuracy/speed/cost tradeoffs. That supports a strong general-purpose score, but the available evidence does not show it as frontier-leading across all areas. | Amazon | 7.4/10 | $0.800 | $3.200 | 300k | Strong multimodal general-purpose model for broad tasks., AWS emphasizes a balance of accuracy, speed, and cost., Supports text, image, and video inputs with text output., Suitable for agentic and RAG-style workflows. | Multimodal assistants, Code generation and editing support, RAG with images/video/documents, Agentic workflow orchestration |
| inclusionAI: Ling 3.1 Flash Listed Oct 2, 2026 Solid frontier-adjacent capability for a flash model, with strong overall intelligence and decent coding benchmarks. Evidence is thinner on strict instruction-following and failure modes. | inclusionAI | 7.3/10 | $0.075 | $0.225 | 262k | strong general reasoning for its tier, good coding performance for a flash model, supports tool use/function calling, long 262K-context availability | general-purpose assistant work, code generation and debugging, tool-using agent workflows, long-document analysis and retrieval |
| Perceptron: Perceptron Mk1.5 Listed Sep 25, 2026 Evidence from the official SDK and package docs supports a capable multimodal embodied-reasoning model, but there are no public benchmark results here and claims are mostly vendor-provided. I’d treat Mk1.5 as strong for grounded perception workflows, with uncertain general-purpose reasoning outside that niche. | Perceptron | 7.3/10 | $0.150 | $1.500 | 37k | Official docs describe strong grounded perception over image, video, and audio, plus tool calling., Supports structured outputs such as boxes, points, polygons, clips, and tracks., Positioned for embodied / physical-AI workflows with temporal clipping and object tracking., SDK and API docs indicate multimodal in-context use with a unified interface. | Robot or agent perception over camera feeds and audio, Object detection, localization, tracking, and OCR with structured outputs, Video question answering and temporal clip extraction, Tool-augmented multimodal workflows for manufacturing, logistics, or security |
| Xiaomi: MiMo-V2.6-Flash Listed Sep 21, 2026 Evidence supports a strong multimodal agent model, but public verification is mostly Xiaomi-authored and recent issues show tool-call repetition and vision bugs. I’m reasonably confident it is capable, but less so on reliability than on raw reasoning/coding. | Xiaomi | 7.3/10 | $0.140 | $0.280 | 1.1M | Official MiMo-V2.6 model card labels it a multimodal, long-context, agent-oriented model with vision, audio, and video-understanding support., The Xiaomi blog and model card show active post-training for tool use / agentic behavior, including | Long-context assistant or agent workflows, Code or tool-using tasks where speed matters, Multimodal Q&A over images, audio, and video, Prototype deployment in OpenAI-compatible serving stacks |
| inclusionAI: Ling 3.0 Flash VL Listed Sep 10, 2026 The creator model card shows a capable open multimodal model with strong context and vision features, but public evidence is narrower than for top frontier models. Benchmarking is mostly maker-provided, so confidence is moderate and coding ability is less directly evidenced. | inclusionAI | 7.3/10 | $0.060 | $0.180 | 131k | Native multimodal model with image and video understanding., Strong long-context support up to 256K tokens., Designed for reasoning, acting, and verification with visual evidence., Public model card reports an AA Intelligence Index score of 42 and benchmarked multimodal behavior. | Vision-language Q&A over images, charts, and documents, Video understanding and event localization, UI understanding and agentic visual task execution, General multimodal assistant workloads |
| Inception: Mercury 2.5 Listed Sep 8, 2026 Inception presents Mercury 2.5 as its most capable production model, with tunable reasoning, parallel tool calls, and schema-aligned JSON, plus strong speed claims. The evidence suggests a capable interactive model, but public substantiation for broad reasoning/coding superiority is still limited. | Inception | 7.3/10 | $0.040 | $0.150 | 260k | Fast diffusion-LLM architecture with very low latency, Supports tunable reasoning, Parallel tool calls and schema-aligned JSON, Strong fit for interactive, latency-sensitive workflows | Low-latency assistants, Search and RAG orchestration, Voice-agent backends, Structured-output agent pipelines |
| Sakana: Fugu Ultra Listed Jun 24, 2026 Fugu Ultra is presented as a learned multi-agent orchestration system rather than a single monolithic model. That makes it interesting for coordination-heavy use cases, but the evidence for broad frontier dominance is limited. | Sakana | 7.3/10 | $5.000 | $30.00 | 1M | Interesting multi-agent architecture, Good for orchestration-heavy tasks, Large context, Potentially strong specialization | Agent orchestration, Workflow automation, Long-context coordination, Research experimentation |
| NVIDIA: Nemotron 3 Ultra Listed Jun 4, 2026 NVIDIA presents Nemotron 3 Ultra as a frontier-oriented orchestration model, but the public evidence is still thinner than for the most established closed flagships. I score it as strong, with slightly lower confidence on instruction following and reliability. | NVIDIA | 7.3/10 | $0.600 | $2.400 | 262k | Strong orchestration and reasoning focus, Large context window, Open frontier-oriented model, Useful for structured workflows | Reasoning and orchestration, Tool-using agents, Structured workflow automation, Long-context tasks |
| NVIDIA: Nemotron 3 Ultra (free) Listed Jun 4, 2026 NVIDIA presents Nemotron 3 Ultra as a frontier-oriented orchestration model, but the public evidence is still thinner than for the most established closed flagships. I score it as strong, with slightly lower confidence on instruction following and reliability. | NVIDIA | 7.3/10 | $0.000 | $0.000 | 1M | Strong orchestration and reasoning focus, Large context window, Open frontier-oriented model, Useful for structured workflows | Reasoning and orchestration, Tool-using agents, Structured workflow automation, Long-context tasks |
| Xiaomi: MiMo-V2.5-Pro Listed Apr 22, 2026 Xiaomi positions MiMo-V2.5-Pro as a flagship agentic coding model with 1M context, but much of the evidence is vendor-reported. I score it as promising but with lower confidence than the best-established frontier models. | Xiaomi | 7.3/10 | $0.435 | $0.870 | 1.1M | Strong agentic and coding focus, Very large context window, Low-cost deployment, Promising benchmark positioning | Software engineering, Long-horizon tasks, Agent workflows, Cost-sensitive deployments |
| Mistral: Mistral Small 4 Listed Mar 16, 2026 Mistral Small 4 is a capable compact model for practical production work, but it still has more limits than the top flagship systems. The evidence suggests a strong small model rather than a frontier leader. | Mistral | 7.3/10 | $0.150 | $0.600 | 262k | compact practical model, good hybrid reasoning and coding, multimodal support, production-friendly cost | light coding tasks, summarization, structured extraction, general production workloads |
| Mistral: Mistral Small 4 (batch) Listed Mar 16, 2026 Mistral Small 4 is a capable compact model for practical production work, but it still has more limits than the top flagship systems. The evidence suggests a strong small model rather than a frontier leader. | Mistral | 7.3/10 | $0.075 | $0.300 | 262k | compact practical model, good hybrid reasoning and coding, multimodal support, production-friendly cost | light coding tasks, summarization, structured extraction, general production workloads |
| ByteDance Seed: Seed 1.6 Listed Dec 23, 2025 Official ByteDance documentation presents Seed1.6 as the family’s general-purpose multimodal model with adaptive thinking and deep reasoning. I score it a bit above the Flash variant on depth, but the public evidence still leaves some uncertainty about exact general capability. | ByteDance Seed | 7.3/10 | $0.250 | $2.000 | 262k | Multimodal general-purpose model, Adaptive deep thinking, GUI and visual understanding, Large 256K context window | Multimodal general assistants, Vision + text reasoning, UI/workflow automation, Long-context problem solving |
| Mistral: Mistral Small 3.2 24B Listed Jun 20, 2025 Official Mistral docs describe Small 3.2 as an update to the Small 3 line, with 128k context and built-in structured outputs/function calling, but Mistral has already deprecated it in favor of Small 4. Good fit for routine instruction-heavy work, though evidence suggests it is not Mistral’s frontier tier. | Mistral | 7.3/10 | $0.094 | $0.250 | 256k | Strong general instruction following for a 24B-class model., Supports structured outputs, function calling, and document Q&A., Multimodal/image input support., Designed for low-latency deployment and everyday assistant tasks. | General-purpose chat and assistant workflows., Structured extraction and tool-using automations., Image-aware support and document Q&A., On-device or cost-sensitive deployments needing good latency. |
| Qwen: Qwen-Plus Listed Feb 1, 2025 Qwen’s public comparison describes Qwen-Plus as competitive with strong proprietary and open models, but still below the very best in some aspects. That supports a strong general-purpose rating, with some uncertainty because the exact deployed variant is not fully transparent. | Qwen | 7.3/10 AI review detailsMedium confidence · Oct 17, 2026 Coding: 6.8/10 Reasoning: 7.6/10 Reliability: 7/10 Instructions: 7.5/10 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $0.260 | $0.780 | 1M | API flagship model with broad general-language focus., Qwen materials compare it against leading models., Appears competitive on general tasks., Likely robust for everyday instruction-heavy use. | General-purpose assistants, Business writing and summarization, Reasoning-heavy chat, API-based production workflows |
| Meta: Llama 3.3 70B Instruct Listed Dec 6, 2024 Meta’s model card supports strong benchmark performance, especially for instruction following and multilingual tasks. It remains a capable open model, but below the best frontier systems on the hardest reasoning problems. | Meta | 7.3/10 | $0.100 | $0.320 | 131k | Strong instruction following, Good multilingual capability, Useful open deployment target, Solid everyday assistant | General assistant tasks, Multilingual chat, Light coding, On-prem deployment |
| OpenAI: GPT-4o (2024-11-20) Listed Nov 20, 2024 OpenAI’s changelog and system card support this snapshot as a refined GPT-4o with modest quality improvements over earlier releases. It remains useful, but it is below later frontier models in this batch. | OpenAI | 7.3/10 | $2.500 | $10.00 | 128k | Good general multimodal utility, Natural writing improvements, Stable structured-output support, Broad applicability | General assistant tasks, Image-assisted workflows, Structured extraction, Creative drafting |
| OpenAI: GPT-4o-mini (batch) Listed Jul 18, 2024 OpenAI’s model docs describe GPT-4o mini as a fast, affordable small model with text and image inputs, structured outputs, and a 128k context window. It is solid for routine production tasks, but still below frontier reasoning models. | OpenAI | 7.3/10 | $0.075 | $0.300 | 128k | Strong small-model generalist, Multimodal text+image input, Structured outputs and tool support, Efficient for high-volume workloads | High-throughput agent backends, Classification and extraction, Vision-enabled lightweight assistants, Cost-sensitive general-purpose automation |
| OpenAI: GPT-4 Listed May 28, 2023 OpenAI’s technical report presents GPT-4 as a multimodal model with human-level performance on several academic benchmarks and improved factuality after alignment. It remains capable, but it is clearly behind the newer reasoning-focused generations, so it should be treated as a strong legacy model rather than a frontier one. | OpenAI | 7.3/10 AI review detailsHigh confidence · Oct 18, 2026 Coding: 7.1/10 Reasoning: 7.4/10 Reliability: 7/10 Instructions: 7.5/10 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $30.00 | $60.00 | 8k | Historically strong multimodal reasoning, Solid instruction-following for its generation, Broadly capable general-purpose model, Still useful as a dependable baseline | Legacy production systems, Multimodal text/image tasks, General-purpose assistance and summarization, Regression baselines for older deployments |
| Dots Studio: Dots3-Note Preview (free) Listed Aug 14, 2026 The dots3-note preview repo describes a broad multimodal MoE with strong long-context and agent-oriented capabilities, which is promising for general utility. Because the evidence is mostly from the vendor repo and release materials, I keep confidence moderate. | Dots Studio | 7.2/10 | $0.000 | $0.000 | 512k | Open-weight multimodal MoE model, Long 512K context window, Supports tool calling and structured outputs, Can process text, images, video, and audio | multimodal assistants, long-context analysis, agent workflows, code-and-document understanding |
| Google: Gemma 4 31B (free) Listed Apr 2, 2026 Official docs show Gemma 4 31B is a 30.7B multimodal open model with 256K context and strong benchmark results, including coding and reasoning. It looks solid for open deployment, but it is still below current flagship proprietary models and is narrower than top closed frontier systems. | 7.2/10 | $0.000 | $0.000 | 262k | Strong multimodal understanding for text and images, Large 256K context window, Native function calling and configurable thinking modes, Solid open-model benchmark results for coding/reasoning at 31B scale | Document and image analysis, Long-context assistant workflows, On-device or self-hosted multimodal apps, General-purpose coding help with vision input | |
| ByteDance Seed: Seed-2.0-Mini Listed Feb 26, 2026 ByteDance’s official Seed2.0 materials present Mini as one of three general-purpose agent models, with multimodal understanding and production-oriented deployment focus. The public evidence is credible but light on independent benchmark detail for this exact mini variant. | ByteDance Seed | 7.2/10 | $0.100 | $0.400 | 262k | General-purpose agent model, Multimodal support, Targets latency-sensitive deployment, Official materials emphasize real-world task handling | Production chat assistants, Agentic task execution, Multimodal Q&A, Cost-sensitive high-concurrency workloads |
| Writer: Palmyra X5 Listed Jan 21, 2026 Writer positions Palmyra X5 as its advanced enterprise agent model with 1M context and built-in tool calling, which supports strong workflow use. Public evidence is mostly vendor-provided, so broad capability claims should be treated cautiously. | Writer | 7.2/10 | $0.600 | $6.000 | 1M | Built for enterprise agent workflows, Large 1M-token context window, Adaptive reasoning focus, Tool-use oriented | Enterprise agent orchestration, Long-document analysis, Workflow automation, Tool-calling assistants |
| Z.ai: GLM 4.6V Listed Dec 8, 2025 GLM-4.6V is presented as a 106B vision-language model with 128K training context and native function calling, aimed at complex multimodal reasoning. Evidence is mostly vendor-provided, so reliability is less well verified. | Z.ai | 7.2/10 AI review detailsMedium confidence · Oct 17, 2026 Coding: 6.8/10 Reasoning: 8/10 Reliability: 6.4/10 Instructions: 6.9/10 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $0.300 | $0.900 | 131k | Strong visual understanding and document/page reasoning, Native multimodal function calling/tool use, Long-context multimodal workflows | Document and screenshot QA, Vision-based agent workflows, Chart/image-grounded extraction and analysis |
| Qwen: Qwen3 Coder Flash Listed Sep 17, 2025 Public Qwen docs present Qwen3-Coder-Flash as the fast model for suggestions, recaps, and tool-use summaries. Strong evidence for coding-focused utility, though public evidence is more ecosystem-oriented than benchmark-heavy. | Qwen | 7.2/10 | $0.195 | $0.975 | 1M | Coding-agent orientation, Tool-use friendly, Optimized for fast/cheap helper loops, Built into Qwen Code workflows | Autonomous programming assistance, Repo edits and bug fixing, Code review helpers, Fast speculative coding suggestions |
| Mistral: Mistral Medium 3.1 Listed Aug 13, 2025 Evidence supports a strong enterprise multimodal generalist with better coding-oriented positioning than a routine assistant, but public benchmark detail for 3.1 is limited. I’m confident it is solid, but not at the very top tier of frontier reasoning. | Mistral | 7.2/10 | $0.400 | $2.000 | 131k | Strong enterprise generalist with multimodal support, Good coding and agentic-use positioning, Native structured-output / tool-use support, Officially updated as an active model version in Aug 2025 | General business assistants, Code generation and editing, Multimodal document/image workflows, Tool-using enterprise agents |
| Mistral: Mistral Medium 3.1 (batch) Listed Aug 13, 2025 This batch variant should inherit the same core capability profile as Mistral Medium 3.1, with the main difference being serving mode rather than intelligence. Public evidence for the exact batch packaging is thin, so I keep confidence moderate. | Mistral | 7.2/10 | $0.200 | $1.000 | 131k | Same underlying model family as Mistral Medium 3.1, Strong enterprise generalist with multimodal support, Good coding and agentic-use positioning, Native structured-output / tool-use support | High-throughput batch summarization, Code transformation pipelines, Multimodal document/image workflows, Tool-using enterprise agents |
| Mistral: Codestral 2508 Listed Aug 1, 2025 Evidence shows Codestral 25.08 is a strong coding-specialist model with solid tooling support and long context. It should be strong for code tasks, but it is not positioned as a top general-purpose reasoner. | Mistral | 7.2/10 | $0.300 | $0.900 | 256k | Purpose-built for code completion and FIM, Supports function calling and structured outputs, Documented long context window, Optimized for low-latency coding workflows | IDE code completion, Fill-in-the-middle editing, Test generation and code correction, Tool-using coding assistants |
| Mistral: Codestral 2508 (batch) Listed Aug 1, 2025 This batch variant is the same Codestral 25.08 model exposed for batch workflows, so capability should match the base model. The main difference is operational, not model quality. | Mistral | 7.2/10 | $0.150 | $0.450 | 256k | Same model family as Codestral 25.08, Supports batching for offline throughput, Purpose-built for code completion and FIM, Supports structured outputs and function calling | Batch code generation jobs, Offline code transformation pipelines, Automated test and snippet generation, Large-scale coding workflow backends |
| Google: Gemini 2.5 Flash Lite Listed Jul 22, 2025 Google positions Flash-Lite as the fastest, lowest-cost Gemini 2.5 model with thinking controls, multimodal I/O, function calling, and structured outputs. Best for lightweight, high-frequency work. | 7.2/10 | $0.100 | $0.400 | 1M | Very fast low-latency responses, Multimodal inputs including image/audio/video/PDF, Function calling and structured outputs, Thinking controls and batch/flex inference | High-volume classification, Simple data extraction, Latency-sensitive assistants, Multimodal lightweight automation | |
| Google: Gemini 2.5 Flash Lite (batch) Listed Jul 22, 2025 Batch-serving variant of Gemini 2.5 Flash-Lite; capability should track the same underlying model. The main difference is operational mode, favoring throughput over interactive depth. | 7.2/10 | $0.050 | $0.200 | 1M | Same core model family as Flash-Lite, Batch-friendly throughput, Multimodal inputs, Function calling and structured outputs | Offline bulk processing, Queued extraction jobs, Large-scale classification pipelines, Batch multimodal preprocessing | |
| Reka Flash 3 Listed Mar 12, 2025 The public Reka Flash 3 model card/config shows a 32K Llama-style instruction model and the official positioning emphasizes chat, coding, instruction-following, and function calling. Evidence is good but not as benchmark-rich as top-tier frontier models, so I rate it as strong but not exceptional. | rekaai | 7.2/10 | $0.100 | $0.200 | 66k | General-purpose instruction-tuned model, Officially positioned for coding and function calling, 32K context in the public config, Reasonably balanced across chat and developer tasks | General chat assistants, Coding help, Function calling, Structured workflow automation |
| Perplexity: Sonar Reasoning Pro Listed Mar 7, 2025 Perplexity documents Sonar as a web-grounded reasoning model, and recent deep-research benchmarks include Sonar Reasoning Pro among leading research systems. Exact base-model details are not fully disclosed, so this is a capability estimate rather than a measured rating. | Perplexity | 7.2/10 | $2.000 | $8.000 | 128k | Web-grounded reasoning workflow, Strong at multi-step search and synthesis, Reasoning-focused presets/controls, Better suited than plain chat models for cited answers | Research Q&A with live sources, Comparative analysis across the web, Factual synthesis with citations, Agentic workflows that benefit from search and reasoning |
| Qwen2.5 72B Instruct Listed Sep 19, 2024 Qwen2.5-72B-Instruct is a strong older flagship open model, but it is well behind the latest frontier releases. | Qwen | 7.2/10 | $0.360 | $0.400 | 33k | Strong instruction tuning, Good multilingual generalization, Useful structured output, Competitive open model | General chat and drafting, Extraction and formatting, Multilingual tasks, Budget deployments |
| Cohere: Command R+ (08-2024) Listed Aug 30, 2024 Cohere frames Command R+ 08-2024 as the stronger option for complex RAG and multi-step tool use. It is highly capable for enterprise retrieval workflows, but specialized rather than universal-best. | Cohere | 7.2/10 | $2.500 | $10.00 | 128k | Strong enterprise RAG utility, Good multi-step tool use, Efficient throughput/latency, Useful for retrieval-heavy tasks | RAG systems, Enterprise assistants, Retrieval and summarization, Tool-using workflows |
| Mistral: Mixtral 8x22B Instruct Listed Apr 17, 2024 Official model-card materials identify this as the instruct-tuned Mixtral 8x22B MoE with 65k context and 64k+ scale. It is a strong open generalist, though still below top frontier models in difficult reasoning and reliability. | Mistral | 7.2/10 | $2.000 | $6.000 | 66k | Strong open MoE generalist, Good balance of quality and efficiency, Longer context than earlier Mixtral versions, Well-supported instruct tuning | General assistant workloads, Coding and technical drafting, Reasoned text generation, Self-hosted high-capacity deployments |
| WizardLM-2 8x22B Listed Apr 16, 2024 Microsoft’s paper and related evaluations support strong open-model capability, but the evidence is older and mostly relative to 2024-era baselines, so I rate it as strong rather than frontier. Confidence is medium because sources are substantive but not exhaustive for the exact 8x22B snapshot. | microsoft | 7.2/10 | $0.620 | $0.620 | 66k | Strong post-training/instruction tuning for chat-style use., Large MoE model with competitive broad-task performance for its era., Good general reasoning and writing quality relative to open-weight peers., Useful for demanding offline/open-model deployments. | General assistant/chat, Reasoning-heavy drafting and analysis, Coding help, Open-weight self-hosted deployment |
| NVIDIA: Nemotron 3.5 Lightning Listed Aug 11, 2026 Evidence supports a capable, efficient agentic model, but public primary sources are thinner than for flagship frontier models. I can assess it as strong for structured/tool-heavy tasks, with less certainty about broad frontier reasoning and coding depth. | NVIDIA | 7.1/10 | $0.070 | $0.200 | 262k | Efficient agentic workflows with high throughput, Strong structured-output and tool-use orientation, Open-weight ecosystem availability and deployment flexibility, Good fit for specialized, task-driven prompting | Tool-using agents, Structured extraction and routing, High-volume batch assistants, Domain-specific workflows where latency matters |
| Cohere: North Mini Code (free) Listed Jun 17, 2026 Cohere’s docs say North Mini Code is a 30B/3B MoE model built specifically for agentic coding, with 256K context and 64K output. It looks strong for software-engineering workflows, but the available evidence is narrower than for large frontier generalists. | Cohere | 7.1/10 | $0.000 | $0.000 | 256k | Purpose-built agentic coding model, 30B total / 3B active MoE efficiency, 256K context and 64K output, Open-source, Apache 2.0, deployable via Cohere | Terminal-based coding agents, Repo-level software engineering tasks, Local or private deployments for code workflows |
| Qwen: Qwen3.5-27B Listed Feb 25, 2026 The exact Qwen3.5-27B model is publicly documented on Hugging Face and referenced in Qwen’s own materials, supporting a competent general-purpose rating. Evidence suggests strong routine capability, but not the top-tier breadth of the largest frontier models. | Qwen | 7.1/10 AI review detailsMedium confidence · Oct 17, 2026 Coding: 6.9/10 Reasoning: 7.2/10 Reliability: 7/10 Instructions: 7.1/10 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $0.260 | $2.600 | 262k | Well-documented Qwen3.5 base model, Strong general-purpose performance for its size, Publicly available on Hugging Face, Useful for multimodal text/image tasks | General chat and Q&A, Moderate-complexity coding, Multimodal assistants, Fine-tuning and deployment experiments |
| Qwen: Qwen3.5-Flash Listed Feb 25, 2026 The official docs show a capable, efficiency-focused multimodal model with broad API features and very long context. I found no comparable public benchmark card for this exact snapshot, so the score is more cautious and reflects likely strong-but-not-flagship capability. | Qwen | 7.1/10 | $0.065 | $0.260 | 1M | Official docs describe native vision-language support with text-image-video input., Supports function calling, structured outputs, and a 1M-token context window., Designed for higher inference efficiency and fast response times., Available as an exact dated snapshot, reducing ambiguity about the version assessed. | Fast multimodal assistant work where latency matters., Tool-using workflows needing structured outputs and web search., High-context summarization or extraction over long mixed-media inputs. |
| Upstage: Solar Pro 3 Listed Jan 27, 2026 Upstage positions Solar Pro 3 as a production-scale enterprise model with stronger reasoning and instruction-following than Solar Pro 2, while preserving serving behavior. Evidence is solid for business deployment and reliability, but less so for top-end coding or frontier general reasoning. | Upstage | 7.1/10 | $0.150 | $0.600 | 131k | Enterprise-oriented production model, Documented gains in reasoning and instruction following, MoE design with stable serving behavior, Compatible upgrade path from Solar Pro 2 | Enterprise chat and assistant workloads, Instruction-following and structured business tasks, Korean/English production use, Reasoning-heavy but efficiency-sensitive applications |
| ByteDance Seed: Seed 1.6 Flash Listed Dec 23, 2025 ByteDance’s official Seed1.6 materials describe a multimodal model family with adaptive thinking, GUI interaction, and deep reasoning. Seed1.6-Flash should be slightly more speed-optimized than the base model, so I rate it strong but not top-tier. | ByteDance Seed | 7.1/10 | $0.075 | $0.300 | 262k | Multimodal understanding, Adaptive thinking/reasoning controls, Large 256K context window, Fast efficiency-oriented variant | Multimodal assistants, GUI or vision-capable workflows, Long-context analysis, Low-latency general assistance |
| Mistral: Ministral 3 14B 2512 Listed Dec 2, 2025 Mistral’s docs describe Ministral 3 14B as the largest Ministral 3 model, with performance comparable to Small 3.2 24B and support for vision, function calling, and structured outputs. It reads as a strong compact general model. | Mistral | 7.1/10 | $0.200 | $0.200 | 262k | Strong small-model text and vision capability, Function calling and structured outputs, Long-context deployment focus | General assistant tasks, Structured extraction and tool use, Local or latency-sensitive deployments |
| Mistral: Ministral 3 8B 2512 Listed Dec 2, 2025 Evidence supports a strong small multimodal model for edge and agentic workflows, but not a frontier generalist. Likely solid on routine reasoning and instruction-following, with expected limits versus larger flagship models. | Mistral | 7.1/10 | $0.150 | $0.150 | 262k | Compact edge model with 256k context and native image understanding., Supports function calling, structured outputs, document Q&A, and batching., Official docs position it as a general-purpose model with agentic/lightweight emphasis., Strong fit for efficient multimodal deployment on constrained hardware. | On-device or edge assistant with image input., Structured extraction and tool-using workflows., Document Q&A over long context inputs., Lightweight multimodal task routing. |
| Mistral: Ministral 3 8B 2512 (batch) Listed Dec 2, 2025 This batch variant is the same capability profile as Ministral 3 8B; the difference is the delivery mode, not the model. Strong small-model utility, but not in the same class as frontier general-purpose systems. | Mistral | 7.1/10 | $0.075 | $0.075 | 262k | Same underlying Ministral 3 8B model as the non-batch variant., Compact multimodal model with 256k context., Supports function calling, structured outputs, and batching., Suitable for long-context, cost-sensitive workloads. | Batch processing of multimodal or document tasks., Long-context extraction jobs., Tool-using agents in offline or queued pipelines., Edge deployment with images and structured outputs. |
| Mistral: Mistral Large 3 2512 Listed Dec 1, 2025 Mistral describes Large 3 as its state-of-the-art open-weight general-purpose multimodal model, and the docs place it in the flagship tier. Evidence is mostly vendor-authored, so I rate it solidly capable but with more uncertainty than the closed-model front-runners. | Mistral | 7.1/10 | $0.500 | $1.500 | 262k | Mistral’s flagship open-weight general-purpose model, Multimodal and function-calling capable, Very cost-efficient relative to larger frontier models, Good for localizable/open-model deployments | Budget-sensitive general chat, Tool-using applications, Multimodal document tasks, Open-model deployment stacks |
| Mistral: Mistral Large 3 2512 (batch) Listed Dec 1, 2025 Mistral describes Large 3 as its state-of-the-art open-weight general-purpose multimodal model, and the docs place it in the flagship tier. Evidence is mostly vendor-authored, so I rate it solidly capable but with more uncertainty than the closed-model front-runners. | Mistral | 7.1/10 | $0.250 | $0.750 | 262k | Mistral’s flagship open-weight general-purpose model, Multimodal and function-calling capable, Very cost-efficient relative to larger frontier models, Good for localizable/open-model deployments | Budget-sensitive general chat, Tool-using applications, Multimodal document tasks, Open-model deployment stacks |
| Amazon: Nova Premier 1.0 Listed Oct 31, 2025 AWS positions Nova Premier as the top-end Nova multimodal model for complex reasoning and agentic workflows, with strong enterprise infrastructure around it. It looks broadly capable, but the evidence base is mostly vendor documentation rather than independent benchmarks. | Amazon | 7.1/10 | $2.500 | $12.50 | 1M | Amazon’s most capable multimodal model for complex tasks., 1M-token context with text, image, and video inputs., Officially positioned for reasoning, agentic workflows, and distillation., Supported through Bedrock Converse/Invoke with enterprise deployment options. | Complex multimodal reasoning in enterprise apps., RAG and agentic workflows over long context., Video/image/text analysis pipelines., Teacher model for distillation and model customization. |
| OpenAI: o3 Mini Listed Jan 31, 2025 OpenAI positions o3-mini as a compact reasoning model optimized for STEM, coding, and science. It is a capable small model, but not a frontier generalist. | OpenAI | 7.1/10 | $1.100 | $4.400 | 200k | Compact STEM reasoning, Reasoning-effort control, Structured outputs, Good cost efficiency | Math and science, Budget coding help, Structured extraction, High-volume reasoning |
| OpenAI: o3 Mini (batch) Listed Jan 31, 2025 OpenAI positions o3-mini as a compact reasoning model optimized for STEM, coding, and science. It is a capable small model, but not a frontier generalist. | OpenAI | 7.1/10 | $0.550 | $2.200 | 200k | Compact STEM reasoning, Reasoning-effort control, Structured outputs, Good cost efficiency | Math and science, Budget coding help, Structured extraction, High-volume reasoning |
| Mistral: Mistral Nemo Listed Jul 19, 2024 Mistral’s official materials present Nemo as a strong 12B-class general model with long context and competitive reasoning/coding for its size. The evidence is decent but still mostly vendor-provided, so I rate it as a strong mid-to-upper-tier small model. | Mistral | 7.1/10 | $0.029 | $0.030 | 131k | Official 12B instruction model with 128K context., Published as a drop-in replacement in the Mistral 7B ecosystem., Claims strong reasoning, world knowledge, and coding for its size., Multilingual focus broadens practical utility. | Long-context summarization, General chat and assistance, Multilingual document tasks, Routine coding and tool workflows |
| OpenAI: GPT-4o Listed May 13, 2024 OpenAI’s docs and system card describe GPT-4o as a versatile flagship with broad multimodal capability. It is still useful, but it is an older snapshot and below later models in this comparison. | OpenAI | 7.1/10 | $2.500 | $10.00 | 128k | Versatile multimodal baseline, Broad tool ecosystem, Good structured-output utility, Stable legacy flagship | General multimodal chat, Image understanding, Structured extraction, Legacy app compatibility |
| OpenAI: GPT-4o (batch) Listed May 13, 2024 OpenAI’s docs and system card describe GPT-4o as a versatile flagship with broad multimodal capability. It is still useful, but it is an older snapshot and below later models in this comparison. | OpenAI | 7.1/10 | $1.250 | $5.000 | 128k | Versatile multimodal baseline, Broad tool ecosystem, Good structured-output utility, Stable legacy flagship | General multimodal chat, Image understanding, Structured extraction, Legacy app compatibility |
| LiquidAI: LFM2.5-24B-A2B Listed Oct 13, 2026 Evidence points to a capable agent-oriented instruct model with built-in tool calling and strong deployment efficiency, but public third-party benchmarks are sparse. Liquid’s own docs also say it is not optimized for coding, so coding capability should be treated as moderate rather than frontier. | LiquidAI | 7.0/10 | $0.050 | $0.700 | 128k | Native tool/function calling with JSON or Pythonic call formatting, Strong fit for agentic workflows and structured outputs, Long context and on-device deployment focus, General-purpose instruct tuning with multilingual support | Tool-using assistants and multi-step agent pipelines, Local/on-prem RAG and document Q&A, Privacy-sensitive customer support workflows, Summarization and general assistant tasks on consumer hardware |
| Xiaomi: MiMo-V2.6-Pro Listed Sep 21, 2026 Xiaomi positions MiMo-V2.6-Pro as its most powerful flagship reasoning model, and its docs plus a third-party benchmark place it high among open models. The evidence is thinner than for the closed frontier models, so the capability estimate is less certain. | Xiaomi | 7.0/10 | $0.435 | $0.870 | 1.1M | Strong open-source multimodal flagship positioning, Large context window and tool support, Good for complex projects and research, Independent benchmark coverage exists, but vendor-reported evidence dominates | Open-weight multimodal applications, Research and cybersecurity workflows, Long-context project work, Cost-sensitive frontier-adjacent deployments |
| DeepSeek: DeepSeek V4 Flash Vision Exp Listed Aug 21, 2026 DeepSeek’s official docs say this legacy multimodal model matched DeepSeek-V4-Flash on text tasks, and the release note says multimodal agent performance improved substantially. It is now retired and routed to V4.1-Flash, so the exact standalone model is no longer current. | DeepSeek | 7.0/10 | $0.216 | $0.647 | 1M | Multimodal image understanding, Matches base Flash text capabilities, Agent/tool-workflow oriented, Officially documented legacy compatibility path | Screenshot and chart analysis, Multimodal agent workflows, General-purpose text tasks, Tool-using automation with images |
| OpenAI: GPT-5 Mini Listed Aug 7, 2025 OpenAI positions GPT-5 Mini as a compact GPT-5 variant for lighter-weight reasoning tasks. It should be useful and efficient, but the available official evidence suggests a clear step down from the flagship models in depth and robustness. | OpenAI | 7.0/10 | $0.250 | $2.000 | 400k | Lower-latency, lower-cost variant, Keeps many GPT-5-era instruction-following improvements, Suitable for high-volume workloads, Supports multimodal input and tool use | Cheap/high-throughput assistants, Routine coding help, Classification and extraction tasks, Fast drafting and summarization |
| Perplexity: Sonar Deep Research Listed Mar 7, 2025 Perplexity describes this as a research-focused model that autonomously searches, reads, and evaluates sources, and recent benchmark papers evaluate it as a deep-research system. Coding is not its main strength, so scores skew toward research and reliability. | Perplexity | 7.0/10 | $2.000 | $8.000 | 128k | Autonomous multi-step retrieval, Strong synthesis across many sources, Designed for complex research tasks, Research-focused rather than generic chat | Deep literature or market research, Multi-source investigation, Evidence gathering and synthesis, Long-form research briefs |
| Cohere: Command R7B (12-2024) Listed Dec 14, 2024 Cohere’s official release presents Command R7B as a small, fast enterprise model with 128K context, optimized for RAG, tool use, and agentic workflows. That supports strong instruction-following and reliability for retrieval-heavy use, with coding more modest than reasoning-focused frontier models. | Cohere | 7.0/10 | $0.037 | $0.150 | 128k | Strong retrieval-augmented generation orientation, Good tool-use and agentic workflow fit, Fast small model with long context, Well-suited for enterprise search and assistants | RAG assistants, Tool-using agents, Enterprise search, Fast conversational help |
| OpenAI: GPT-4o (2024-08-06) Listed Aug 6, 2024 OpenAI’s docs and system card support this earlier GPT-4o snapshot as a capable general multimodal model. It is solid, but older and clearly behind the strongest current models in this comparison. | OpenAI | 7.0/10 | $2.500 | $10.00 | 128k | Broad multimodal capability, Structured outputs support, Good general-purpose utility, Stable older snapshot | Legacy GPT-4o deployments, Image + text tasks, Structured JSON extraction, General chat and drafting |
| Meta: Llama 3.1 8B Instruct Listed Jul 23, 2024 Meta’s official model card shows a solid step up from Llama 3 8B across instruction, code, math, tool-use, and multilingual benchmarks, but it remains an 8B model with clear ceiling effects. Good small generalist, not a frontier model. | Meta | 7.0/10 AI review detailsHigh confidence · Oct 17, 2026 Coding: 7/10 Reasoning: 6.8/10 Reliability: 6.8/10 Instructions: 7.2/10 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $0.050 | $0.080 | 131k | Strong open 8B instruct model with broad community support., Good tool-use and instruction-following relative to size., Official benchmark tables show clear gains over Llama 3 8B on several tasks., Large 128K-context variant improves long-input utility. | Local assistant deployment, Light coding help, Summarization and chat, Multilingual routine tasks |
| Upstage: Solar Mini 4 Listed Sep 23, 2026 Evidence supports a competent small-to-mid model with agentic features, but I found no public benchmark set for Solar Mini 4 itself. Scores are therefore a cautious estimate based on official API capabilities and Solar Mini family positioning, not measured head-to-head results. | Upstage | 6.9/10 | $0.050 | $0.200 | 524k | Official Upstage docs expose `solar-mini4` and `solar-mini4-260922` as current API models, with full, reasoning-effort control from `none` to `max` and support for structured outputs., Upstage positions Solar Mini as a small, fast model for lightweight reasoning and cost-efficient, tasks, suggesting a practical general-purpose assistant profile rather than a niche-only model. | Agentic workflows that need tool calling and JSON/structured output., Lightweight business assistants, summaries, and form-filling tasks., Korean/English enterprise text tasks where cost and speed matter., Applications that benefit from controllable reasoning levels without moving to a larger flagship. |
| IBM: Granite 4.2 8B Listed Aug 31, 2026 IBM’s public materials support Granite 4.2 8B as a balanced reasoning model with solid benchmark coverage for its size. It looks competent and efficient, but not close to the biggest frontier systems on raw capability. | IBM | 6.9/10 | $0.060 | $0.250 | 131k | Efficient small model, Balanced enterprise reasoning, Multilingual utility, Good for constrained deployments | On-device or efficient server deployment, Enterprise assistants, Multilingual dialogue, Routine coding and math tasks |
| Sakana: Sakana Namazu Listed Aug 11, 2026 The available sources support Namazu as a Japanese-focused, reasoning-oriented model with strong instruction handling and tool support. Public evidence is thinner on coding breadth, so I would treat it as a strong localized assistant rather than a general coding leader. | Sakana | 6.9/10 | $0.950 | $4.000 | 262k | Japanese-specialized instruction following, Extended thinking support, Multimodal input support, Tool calling in the API | Japanese business assistants, Japanese Q&A and drafting, Tool-using support workflows, Localized enterprise chat |
| Google: Gemini 3.1 Flash Lite Listed May 7, 2026 Google’s official docs present Gemini 3.1 Flash-Lite as a low-latency multimodal model for high-frequency tasks. That fits a competent everyday model, but the docs frame it as efficiency-first rather than top-end reasoning or coding. | 6.9/10 AI review detailsHigh confidence · Oct 17, 2026 Coding: 6.4/10 Reasoning: 6.9/10 Reliability: 7.3/10 Instructions: 7.4/10 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $0.250 | $1.500 | 1M | Officially multimodal across text, image, video, audio, and PDF, Optimized for low latency and high-volume workloads, Supports function calling and structured outputs, Clear Google model docs with exact stable model ID | High-throughput assistants, Lightweight multimodal extraction, Fast summarization and translation, Tool-using agent workflows with cost/latency constraints | |
| Google: Gemini 3.1 Flash Lite (batch) Listed May 7, 2026 This batch alias points to the same Gemini 3.1 Flash-Lite model family described in Google’s docs. Because the underlying model is documented as efficiency-oriented, the capability profile should match the non-batch version closely. | 6.9/10 AI review detailsHigh confidence · Oct 17, 2026 Coding: 6.4/10 Reasoning: 6.9/10 Reliability: 7.3/10 Instructions: 7.4/10 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $0.125 | $0.750 | 1M | Same underlying Gemini 3.1 Flash-Lite model family, Batch-friendly for high-volume workloads, Official docs indicate multimodal and structured-output support, Clear stable model naming in Google docs | Bulk multimodal processing, Batch extraction and classification, High-volume assistant pipelines, Cost-sensitive agent backends | |
| Google: Gemma 4 26B A4B Listed Apr 3, 2026 Gemma 4 26B A4B IT looks strong for its class, but it is still below larger frontier models on open-ended tasks. I score it as a competent, efficient general-purpose model rather than a top flagship. | 6.9/10 | $0.068 | $0.225 | 262k | Strong for its size, Efficient MoE design, Good reasoning/coding for class, Low-cost deployment | Cost-sensitive apps, On-device or efficient inference, Basic reasoning, Moderate coding help | |
| Google: Gemma 4 26B A4B (free) Listed Apr 3, 2026 Gemma 4 26B A4B IT looks strong for its class, but it is still below larger frontier models on open-ended tasks. I score it as a competent, efficient general-purpose model rather than a top flagship. | 6.9/10 | $0.000 | $0.000 | 262k | Strong for its size, Efficient MoE design, Good reasoning/coding for class, Low-cost deployment | Cost-sensitive apps, On-device or efficient inference, Basic reasoning, Moderate coding help | |
| MiniMax: MiniMax M2-her Listed Jan 23, 2026 Evidence supports a dialogue-first model tuned for roleplay and long-turn consistency, with public MiniMax material claiming strong 100-turn performance. There is little primary evidence of broad coding strength, so I score it as specialized rather than general-purpose. | MiniMax | 6.9/10 | $0.300 | $1.200 | 66k | Strong multi-turn character consistency, Roleplay and narrative continuity, Good tone and persona control, Designed for immersive dialogue | Character chat and roleplay, Interactive fiction / story companions, Persona-driven assistants, Long conversational scenarios |
| Amazon: Nova 2 Lite Listed Dec 2, 2025 AWS documents Nova 2 Lite as a cost-efficient multimodal model for text, images, and video with a 1M-token context window and enterprise use cases. The public evidence supports a solid general-purpose multimodal model. | Amazon | 6.9/10 | $0.300 | $2.500 | 1M | Multimodal text-image-video input support, Very large 1M-token context window, Designed for fast, cost-efficient enterprise automation | Document processing, Customer support automation, Long-context multimodal summarization |
| Z.ai: GLM 4.5V Listed Aug 11, 2025 The evidence points to a capable vision-language model with particularly strong multimodal reasoning and agentic use cases. I rate coding and reliability below the top text-only frontier because the public evidence is centered on VLM benchmarks and product claims, not broad text-agent depth. | Z.ai | 6.9/10 | $0.600 | $1.800 | 66k | Vision-language model aimed at multimodal reasoning, Officially described as strong on image, video, document, and GUI-agent tasks, Large MoE backbone with substantial active capacity, Open model card with paper and repository references | Image and video understanding, Document parsing and visual QA, GUI automation / multimodal agents, Multimodal reasoning research |
| Cohere: Command R (08-2024) Listed Aug 30, 2024 Cohere’s docs identify the August 2024 update as improving RAG, tool use, math, code, and reasoning, but the public evidence is more product-oriented than benchmark-dense. That supports a strong workflow model with some uncertainty about exact standalone reasoning depth. | Cohere | 6.9/10 | $0.150 | $0.600 | 128k | Optimized for retrieval-augmented generation and tool use., Strong multilingual enterprise-oriented behavior., Good instruction following for workflow integration., Useful for grounded answers with external context. | RAG assistants, Enterprise search and grounding, Tool-using agents, Multilingual customer support |
| OpenAI: GPT-4 Turbo Listed Apr 9, 2024 GPT-4 Turbo is capable but older, sitting below later GPT-4o/5/6-era models. | OpenAI | 6.9/10 | $10.00 | $30.00 | 128k | Good vision support, Reasonably strong coding, Useful tool calling, Still adequate for many workflows | Legacy app support, Vision-enabled assistants, General drafting, Moderate-complexity coding |
| OpenAI: GPT-4 Turbo (batch) Listed Apr 9, 2024 GPT-4 Turbo is capable but older, sitting below later GPT-4o/5/6-era models. | OpenAI | 6.9/10 | $5.000 | $15.00 | 128k | Good vision support, Reasonably strong coding, Useful tool calling, Still adequate for many workflows | Legacy app support, Vision-enabled assistants, General drafting, Moderate-complexity coding |
| Qwen: Qwen3.8 Omni Flash Listed Sep 21, 2026 Official Qwen docs position this as a next-gen omni-modal model with strong audio/video and multimodal features, but the public evidence is mostly vendor-provided and not a broad third-party benchmark set. Reasoning appears solid and multimodal breadth is strong, while coding and reliability are less clearly evidenced than in text-first frontier Ls | Qwen | 6.8/10 | $0.150 | $0.470 | 1M | native text, image, audio, and video input support, audio-video understanding and summarization, agent-like audio function calling / tool-use workflows, strong multilingual omni-modal coverage | meeting and lecture summarization from audio/video, audio-visual QA and scene understanding, multimodal assistant workflows with tool calling, OCR, image QA, and video description tasks |
| Qwen: Qwen3 Coder Next Listed Feb 4, 2026 The strongest evidence says Qwen3-Coder-Next is an experimental coding model tuned for agentic coding and local development, with coding-specific training and tool support. That supports a high coding score, but the available evidence is narrower for general reasoning and broad instruction robustness. | Qwen | 6.8/10 | $0.120 | $0.800 | 262k | Explicitly optimized for coding agents and local development, Supports function calling and structured output, Long context support for repo-scale workflows, Open-weight and available in coding-focused tooling | Coding agents and terminal-based development, Repo refactors and bug fixing, Environment-interaction and tool-driven code tasks, Local developer workflows with structured outputs |
| OpenAI: GPT-5 Nano (batch) Listed Aug 7, 2025 OpenAI documents gpt-5-nano as the fastest, cheapest GPT-5 variant and says it is best for summarization/classification, supporting strong small-model utility but not frontier-depth reasoning or coding. | OpenAI | 6.8/10 | $0.025 | $0.200 | 400k | Very fast, low-cost GPT-5 variant with a 400K context window., Supports text and image input plus reasoning tokens., Docs position it for summarization and classification., Shares GPT-5-era instruction-following improvements. | High-throughput summarization and classification., Lightweight multimodal assistants., Latency-sensitive tool workflows., Drafting, extraction, and routing tasks. |
| Perplexity: Sonar Pro Listed Mar 7, 2025 Perplexity positions Sonar API as direct web-search-powered LLM responses, and its docs frame it as the simplest search+answer path. Public evidence is stronger on the platform workflow than on exact standalone benchmark behavior, so this is a cautious estimate. | Perplexity | 6.8/10 | $3.000 | $15.00 | 200k | Web search plus answer synthesis, Good fit for source-backed responses, Long-context web research use cases, Built-in search-centered API workflow | General research and fact-finding, Source-backed summaries, Current-events question answering, Simple research agents that need grounded outputs |
| Z.ai: GLM 5.3 Flash Listed Aug 26, 2026 GLM-5.3-Flash appears capable and efficient, with official/model-card evidence for native multimodality and a large parameterized release. Still, the strongest evidence is thinner and more vendor-controlled than for the top OpenAI/Anthropic models, so I rate it as competent rather than frontier-leading. | Z.ai | 6.7/10 | $0.048 | $0.200 | 1M | Efficient and cost-effective, Native multimodal support, Good fit for long-context workflows, Appears competitive for coding at its tier | High-throughput assistant workloads, Budget-conscious coding help, Long-context summarization, Multimodal extraction and simple agent tasks |
| Z.ai: GLM 5.3 Flash (batch) Listed Aug 26, 2026 GLM-5.3-Flash appears capable and efficient, with official/model-card evidence for native multimodality and a large parameterized release. Still, the strongest evidence is thinner and more vendor-controlled than for the top OpenAI/Anthropic models, so I rate it as competent rather than frontier-leading. | Z.ai | 6.7/10 | $0.050 | $0.180 | 1M | Efficient and cost-effective, Native multimodal support, Good fit for long-context workflows, Appears competitive for coding at its tier | High-throughput assistant workloads, Budget-conscious coding help, Long-context summarization, Multimodal extraction and simple agent tasks |
| Meituan: LongCat 2.0 Listed Jul 20, 2026 Promising for long-context coding and agentic work, but the public evidence is mostly vendor-provided and thin on independent benchmarks. | Meituan | 6.7/10 AI review detailsLow confidence · Oct 18, 2026 Coding: 7.4/10 Reasoning: 6.8/10 Reliability: 5.7/10 Instructions: 6.2/10 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $0.300 | $1.200 | 1M | Sparse-MoE design with very large total scale and ~48B active parameters per token, Strong focus on long-context, repository-level edits, and agentic workflows, Built and evaluated for coding-oriented harnesses and tool-heavy workflows, Open-source release with model card and implementation details | Repository-level code assistance, Long-horizon agent workflows, Tool-using coding tasks, Long-context document or codebase analysis |
| NVIDIA: Nemotron 3 Nano Omni (free) Listed Apr 28, 2026 NVIDIA’s report and deployment docs describe a strong multimodal perception model, but it is positioned as a sub-agent rather than a top generalist. That supports solid multimodal utility, with more limited evidence for top-tier coding or broad reasoning. | NVIDIA | 6.7/10 | $0.000 | $0.000 | 256k | Native audio, image, and video support, Designed for document understanding and long multimodal context, Optimized for tool-calling and agent systems, Official NVIDIA docs and report describe exact checkpoint and deployment paths | Multimodal perception sub-agent, Video/audio understanding pipelines, Document extraction and QA, Tool-using enterprise assistants |
| Qwen: Qwen3.5-9B Listed Mar 10, 2026 Qwen’s official ecosystem identifies Qwen3.5 as a multimodal foundation family, and the 9B checkpoint is a compact member of that line. That suggests solid routine capability, but at 9B parameters it is unlikely to match the strongest frontier models in hard reasoning or robustness. | Qwen | 6.7/10 | $0.100 | $0.150 | 262k | Multimodal foundation model, Efficient 9B-class footprint, Supports reasoning and tool use, Long context and vision-language design | Lightweight multimodal assistants, General chat and instruction following, Reasoning on routine tasks, Developer prototyping with smaller models |
| Google: Gemini 3.1 Flash Lite Preview Listed Mar 3, 2026 Google’s docs position Gemini 3.1 Flash-Lite Preview as a low-latency, cost-effective multimodal model for lightweight tasks, with broad API capabilities. Evidence supports strong utility and reliability for routine workloads, but not frontier depth versus larger Gemini tiers. | 6.7/10 | $0.250 | $1.500 | 1M | Multimodal inputs including audio, video, image, and PDF, Large context and structured-output support, Designed for high-frequency, lightweight tasks, Thinking, function calling, code execution, and search grounding supported | High-volume classification and extraction, Translation and transcription pipelines, Document triage and summarization, Low-cost routing/classifier steps in agent systems | |
| OpenAI: GPT-5 Image Listed Oct 14, 2025 The official OpenAI materials describe a notably stronger image system with better instruction following, reasoning/tool use, and denser text rendering. Because the public docs are about the broader Images 2.0 stack rather than a standalone GPT-5 Image model card, I keep the score solid but not maximal. | OpenAI | 6.7/10 | $10.00 | $10.00 | 400k | Strong image-generation quality with contextual understanding, Good instruction following for image editing and composition, Reasoning/tool-use support in the broader Images 2.0 stack, Designed for dense text and more complex prompts | High-quality image generation, Conversational image editing, Complex art direction and layout prompts, Workflows needing reasoning-guided image synthesis |
| Qwen: Qwen2.5 VL 72B Instruct Listed Feb 1, 2025 Qwen2.5-VL is positioned as a strong vision-language model with documented gains on image and video understanding, including over earlier Qwen-VL versions. That supports high multimodal capability, though not coding specialization. | Qwen | 6.7/10 | $0.800 | $1.000 | 128k | Strong multimodal understanding across images, charts, layouts, and documents., Official materials describe broad visual-language competence., Reported improvements over prior Qwen-VL generations., Good fit for visually grounded instruction following. | Image/document understanding, Chart and layout analysis, Visual question answering, Multimodal assistants |
| MiniMax: MiniMax-01 Listed Jan 15, 2025 MiniMax’s paper and official open-source announcement describe a 456B MoE text model with 45.9B activated parameters and a companion vision-language model, emphasizing long-context efficiency. That suggests strong capability on long-context comprehension, but exact public evidence for coding and broad instruction-following is thinner, so confidence | MiniMax | 6.7/10 | $0.200 | $1.100 | 1M | Very long-context architecture, Open-sourced text and vision-language variants, Strong fit for document-heavy workflows, Innovation in attention efficiency | Long-document analysis, Multimodal understanding, Agentic workflows with large context, Research experimentation |
| Meta: Llama 3.1 70B Instruct Listed Jul 23, 2024 A strong open-weight generalist, but below the 405B variant and top proprietary frontier models overall. Reliable and useful, though clearly not a current flagship. | Meta | 6.7/10 | $0.400 | $0.400 | 131k | Strong open-weight generalist, Good dialogue quality, Large context for its class, Well-documented model family | Open-weight deployment, General chat, Instruction following, Light-to-moderate coding |
| Qwen: Qwen3 30B A3B Listed Apr 28, 2025 Public evidence shows Qwen3-30B-A3B is a strong small MoE model and a clear step up from older mid-sized open models. However, the available evidence is mostly self-reported and does not justify calling it frontier-level. | Qwen | 6.6/10 | $0.120 | $0.500 | 41k | Good reasoning for its size, Efficient MoE architecture, Competitive open-model baseline for agents, Useful long-context support relative to class | Budget-conscious assistants, Open-weight prototyping, Moderate-complexity reasoning tasks, Multilingual or self-hosted deployments |
| Meta: Llama 4 Scout Listed Apr 5, 2025 Meta’s docs describe Llama 4 Scout as a natively multimodal model with very large context and efficient deployment, supporting broad vision-language utility. | Meta | 6.6/10 | $0.100 | $0.300 | 1.3M | Native multimodal model with image and text understanding., Meta describes Scout as efficient enough for a single H100 GPU., Large context window supports long-document workflows., Positioned as a strong general-purpose multimodal model. | Vision-language assistants., Long-context document analysis., General-purpose chat and summarization., Moderate-complexity agentic workflows. |
| Google: Gemma 3 12B Listed Mar 13, 2025 Google’s Gemma 3 card shows the 12B instruct model is stronger than 4B across reasoning and coding benchmarks while keeping multimodal and long-context design. | 6.6/10 | $0.050 | $0.150 | 131k | Multimodal Gemma 3 model with stronger published benchmark results than 4B., Good balance of capability and deployability for local/self-hosted use., Large 128K context for the 4B/12B/27B sizes., Solid text, image, QA, summarization, and reasoning coverage. | General-purpose local assistant., Image + text analysis., Coding help and routine reasoning tasks., Enterprise prototypes needing a smaller open model. | |
| Nous: Hermes 3 70B Instruct Listed Aug 18, 2024 Hermes 3 is a capable open instruction model with emphasis on function calling and structured output. The evidence here is mostly model-card level, so I keep the overall score moderate. | Nous | 6.6/10 | $0.700 | $0.700 | 131k | Good open instruction model, Structured output support, Useful agentic tuning, Decent multi-turn conversation | General open-weight chat, Structured extraction, Light coding help, Roleplay/conversation |
| Anthropic: Claude Haiku 5.5 (batch) Listed Oct 7, 2026 Anthropic’s launch material says Haiku 5.5 is the fastest, cheapest small Claude, with improved safety and solid benchmark gains over Haiku 4.5. Evidence is vendor-led and current; independent third-party validation for this exact model is limited, so scores stay moderate. | Anthropic | 6.5/10 | $0.050 | $0.250 | 1M | Fast, low-cost model for high-volume routine work, Strong fit for summarization, compaction, classification, and database-query style tasks, Designed to pair with larger Claude models as a subagent for coding and agent workflows, Supports browser use and is positioned as the fastest Claude model at standard speed | Bulk summarization and document compaction, Routing, classification, and extraction pipelines, Subagent support in coding or agent systems, Low-latency customer support and browser-based workflows |
| Google: Gemma 2 27B Listed Jul 13, 2024 Google’s Gemma 2 materials describe the 27B variant as a text-to-text open model suitable for question answering, summarization, and reasoning. It is a capable open model, but the available evidence does not justify scoring it near frontier closed models. | 6.5/10 | $0.650 | $0.650 | 8k | Open-weight generalist, Solid text generation and reasoning for its size, Good deployment flexibility, Well-documented family and model card | Local or self-hosted assistants, General text generation, Summarization and extraction, Moderate-complexity reasoning tasks | |
| Inception: Mercury 2 Listed Mar 4, 2026 Official sources describe Mercury 2 as a fast reasoning diffusion model with tool use and JSON output, but public evidence is concentrated on speed and production fit rather than broad benchmark leadership. Good for latency-sensitive agentic apps; less evidence for top-tier general coding/reasoning. | Inception | 6.4/10 | $0.250 | $0.750 | 128k | Very low-latency diffusion-style generation, Native tool use and schema-aligned JSON output, Good fit for agent loops and retrieval pipelines, Strong speed/throughput focus | Real-time voice or chat agents, High-volume extraction and summarization, Tool-using workflows where latency matters, Lightweight RAG pipelines |
| Nous: Hermes 3 405B Instruct Listed Aug 16, 2024 The Hermes 3 report and model card describe a generalist instruct/tool-use model with strong reasoning, creative abilities, and improved function calling. Evidence is mostly self-reported and older, so it looks competent and flexible, but not clearly in the top frontier tier overall. | Nous | 6.3/10 | $1.000 | $1.000 | 131k | Strong open-weight generalist positioning, Useful function calling and structured output support, Good long-context conversational design, More steerable than many base open models | Open-weight assistant baselines, Structured extraction and JSON-heavy tasks, Roleplay/conversation experiments, Custom fine-tuning and local deployment |
| Pareto 26.9 Listed Oct 13, 2026 Evidence is mostly vendor/OpenRouter catalog material, so capability claims are not independently benchmark-verified. It appears competent for multimodal, coding, and agentic tasks, but frontier-level performance and reliability remain unproven. | unbiased | 6.2/10 | $2.500 | $7.500 | 262k | Supports text and image input, Built for research, coding, and agentic workflows, Large 262K context window, Tool calling and JSON-style structured output support are advertised | Multimodal research assistance, Code generation and debugging support, Long-context document analysis, Tool-using agent workflows |
| Magnum v4 72B Listed Oct 22, 2024 The public evidence shows Magnum v4 72B is a Qwen2.5-72B-derived fine-tune, but not enough rigorous, model-specific evaluation to justify a frontier score. It looks strongest as a prose-oriented model, with unclear gains outside writing. | anthracite-org | 6.2/10 | $2.500 | $5.000 | 33k | A Qwen2.5-72B-based fine-tune aimed at prose quality., Likely strong at stylistic writing and formatted outputs., Large parameter scale relative to many hobbyist fine-tunes., May be useful for polished long-form generation. | Long-form prose generation, Style imitation and rewrite tasks, Structured creative writing, Dialogue-heavy drafting |
| Perplexity: Sonar Pro Search Listed Oct 30, 2025 Perplexity’s Sonar Pro Search is clearly optimized for research, synthesis, and citations rather than general coding or pure reasoning breadth. Reliability is comparatively strong because the product is explicitly web-grounded and citation-forward, but it is still a search system, not a general frontier model. | Perplexity | 6.0/10 | $3.000 | $15.00 | 200k | Built for deep web research with multi-step search and synthesis., Produces cited answers grounded in live sources., Pro Search supports model selection and complex query handling., Strong transparency and verification story relative to chat-only models. | Research-heavy question answering., Current-events and source-grounded analysis., Literature, market, and policy scanning., Citation-backed answer generation. |
| Morph: Morph V3 Large Listed Jul 7, 2025 Morph V3 Large is clearly specialized for code application rather than general language reasoning, and the best evidence is largely vendor-reported. It looks excellent for edit-apply workflows, but its general capability is limited by design. | Morph | 6.0/10 | $0.900 | $1.900 | 262k | Specialized for complex code edits, High throughput and long context, Strong reported merge accuracy for semantic edits, Purpose-built apply model rather than general chat model | Applying diffs to codebases, Complex multi-file code transformations, Editing pipelines in coding agents, Rapid semantic merge of generated edits |
| OpenAI: GPT-4.1 Nano Listed Apr 14, 2025 OpenAI says GPT-4.1 nano is the fastest and most cost-efficient GPT-4.1 variant, optimized for instruction following and tool calling with low latency. That makes it useful for simple, high-throughput work, but the official guidance explicitly recommends GPT-5 nano for more complex tasks. | OpenAI | 6.0/10 | $0.100 | $0.400 | 1M | Very fast and cost-efficient., Still supports tool calling, structured outputs, and vision input., 1M-token context window., Well-suited to simple instruction-following tasks. | Classification and routing., High-volume extraction and tagging., Cheap assistant backends for straightforward prompts., Multimodal triage where latency matters most. |
| OpenAI: GPT-4.1 Nano (batch) Listed Apr 14, 2025 This batch variant should be evaluated as GPT-4.1 nano with batch serving, not as a different capability tier. Official docs still frame it as best for simple speed-sensitive workloads rather than deep reasoning or coding. | OpenAI | 6.0/10 | $0.050 | $0.200 | 1M | Same underlying model as GPT-4.1 nano., Very fast and cheap., Tool calling, structured outputs, and vision input supported., Good for simple, repetitive tasks at scale. | Batch classification and extraction., Routing and lightweight agent steps., Cheap multimodal triage., High-volume summarization of short inputs. |
| Perplexity: Sonar Listed Jan 27, 2025 Perplexity’s official docs frame Sonar as a web-grounded, citation-focused model/API and note its transition toward the Agent API; that supports high reliability for search tasks but not evidence of frontier coding or broad reasoning. Public material is sparse on exact general benchmarks, so confidence is limited. | Perplexity | 6.0/10 | $1.000 | $1.000 | 127k | Web-grounded answers with citations, Optimized for search/retrieval-centric workflows, Designed for fast, production API use, Good fit for answer-engine style responses | Cited web Q&A, Research assistants, Search-heavy customer support, Source-grounded summaries |
| Amazon: Nova Micro 1.0 Listed Dec 5, 2024 AWS positions Nova Micro as the fastest, lowest-cost text model in the Nova family, optimized for latency-sensitive workflows rather than top-end capability. Evidence supports solid routine performance, but not frontier reasoning or coding depth. | Amazon | 6.0/10 | $0.035 | $0.140 | 128k | Very low-latency text-only model for short-turn interactions., Good for classification, routing, extraction, and summarization., AWS docs position it for latency-sensitive, low-cost deployment., Supports long context for text workflows. | Chatbots and support triage, Document summarization and extraction, Routing/classification pipelines, Lightweight agent tool orchestration |
| LiquidAI: LFM2.5-2.6B (free) Listed Aug 11, 2026 Evidence supports a capable compact agent model for planning and tool use, but the public evidence does not show frontier coding strength. Liquid itself positions it for agentic workflows and warns away from harder agentic coding tasks. | LiquidAI | 5.9/10 AI review detailsMedium confidence · Oct 17, 2026 Coding: 4.8/10 Reasoning: 6.6/10 Reliability: 5.8/10 Instructions: 6.2/10 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $0.000 | $0.000 | 66k | Compact agentic workflows, Tool use and planning, Structured outputs, On-device deployment | Lightweight data extraction, Simple RAG assistants, Local multi-step task automation, Form or document structuring |
| AionLabs: Aion-3.0 Listed Jul 7, 2026 Aion Labs documents Aion 3.0 as a GLM-based collaborative multi-model system aimed at stronger narrative structure and nuanced darker themes. The evidence supports a storytelling-oriented capability profile, but there is little public benchmark evidence for broad coding or general reliability. | AionLabs | 5.9/10 | $3.000 | $6.000 | 131k | Built as a multi-model system on the GLM family of models., Official docs say multiple specialized models contribute to each response., Docs claim stronger narrative structure and better tension/conflict handling than earlier variants., Reasoning is supported and configurable in the API. | Roleplay and storytelling with more structured arcs., Creative drafting where scene-level balance and conflict matter., Assistant tasks that benefit from a reasoning-enabled, tool-capable API. |
| Google: Nano Banana Pro (Gemini 3 Pro Image Preview) Listed Nov 20, 2025 Official Google sources describe a very capable image-generation/editing model with strong grounding and text rendering. Coding is largely not the point here, so that score stays low despite strong visual and reasoning-oriented capabilities. | 5.9/10 | $2.000 | $12.00 | 66k | Advanced image generation/editing model built on Gemini 3 Pro., Strong emphasis on world knowledge, grounding, and improved text rendering., Supports text, images, and iterative visual editing workflows., Useful for infographic, poster, and diagram generation. | High-fidelity image generation and editing., Infographics and diagram creation from prompts., Image-based creative iteration with text rendering., Visualizing real-world or current-information concepts. | |
| IBM: Granite 4.0 Micro Listed Oct 20, 2025 IBM documents Granite 4.0 H Micro as a 3B hybrid dense model designed for low latency and as a building block for tasks like function calling. That points to a competent utility model, not high-end general intelligence. | IBM | 5.9/10 | $0.017 | $0.112 | 131k | Small, low-latency model, Built for structured and long-context tasks, Useful as an agentic building block | On-device or low-latency assistants, Function-calling subroutines, Structured extraction and routing |
| Qwen: Qwen3 VL 8B Instruct Listed Oct 14, 2025 This looks like a competent small VLM, but I found only sparse exact-model evidence beyond the repository metadata, so I’m cautious. It should work well for routine multimodal tasks, but I would not expect frontier-level reasoning or robustness from an 8B-class model. | Qwen | 5.9/10 | $0.250 | $0.750 | 262k | Multimodal text-image-video instruction model, Likely efficient and easy to deploy relative to larger VLMs, Part of the newer Qwen3-VL series, Useful for general visual understanding tasks | Lightweight image captioning, Document and screenshot QA, Prototype multimodal assistants, Video-frame analysis on constrained hardware |
| Sao10K: Llama 3.3 Euryale 70B Listed Dec 18, 2024 The available public source strongly indicates this is Sao10K’s roleplay-focused Euryale fine-tune on the Llama 3.3 70B family, but there is little substantive benchmark evidence for exact general-purpose capability. I therefore treat it as a specialized chat model rather than a broadly elite generalist. | Sao10K | 5.9/10 AI review detailsLow confidence · Oct 17, 2026 Coding: 5.6/10 Reasoning: 6/10 Reliability: 5.4/10 Instructions: 6.3/10 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $0.650 | $0.750 | 131k | Large 70B-class Llama-based fine-tune, Likely strong conversational fluency, Long context support in the family lineage, Specialized tuning can help style-sensitive chat | Creative roleplay, Character chat, Style-heavy conversational writing, General chat when persona control matters |
| Sao10K: Llama 3.1 Euryale 70B v2.2 Listed Aug 28, 2024 The available evidence consistently describes Euryale v2.2 as a Sao10K fine-tune focused on creative roleplay, not a general-purpose frontier assistant. It has a 131K context window and inherits a strong base, but there is limited public, exact-model benchmark evidence for coding or reliable instruction-following, so those scores stay moderate. | Sao10K | 5.9/10 | $0.850 | $0.850 | 131k | Creative roleplay specialization, Large 70B-class model with 131K context, Useful for style, persona, and dialogue consistency, Widely deployed as an open fine-tune of Llama 3.1 70B | creative roleplay and character chat, persona-driven writing, dialogue generation with long context, light general assistant work when style matters more than precision |
| OpenAI: GPT-3.5 Turbo Listed May 28, 2023 OpenAI documents GPT-3.5 Turbo as a legacy chat/non-chat model with 16,385 context and 2021 cutoff. It remains useful for routine tasks, but OpenAI now recommends GPT-4o mini instead for better capability and speed. | OpenAI | 5.9/10 | $0.500 | $1.500 | 16k | Solid baseline chat and completion model, Supports function calling and structured outputs, Large ecosystem support and well-documented behavior, Useful for routine text/code tasks | Simple chat assistants, Extraction and structured JSON generation, Basic coding help and templating |
| OpenAI: GPT-3.5 Turbo (batch) Listed May 28, 2023 This appears to be the batch-served variant of GPT-3.5 Turbo, so capability should match the base model rather than add new intelligence. Batch changes delivery mode, not the underlying model quality. | OpenAI | 5.9/10 | $0.250 | $0.750 | 16k | Same underlying GPT-3.5 Turbo model family, Works for batch-oriented text workflows, Supports structured outputs and function calling | Bulk extraction jobs, Offline batch summarization, Low-cost legacy text generation |
| Mistral: Voxtral Small 24B 2507 Listed Oct 30, 2025 The official Mistral Voxtral materials position Voxtral Small as an audio-first model for transcription, translation, and audio understanding, with text performance retained. Evidence for broad reasoning or coding strength is limited. | Mistral | 5.8/10 | $0.100 | $0.300 | 33k | Strong audio input/transcription focus, Audio understanding and translation, Useful for speech-centric workflows | Speech transcription, Audio translation, Voice-note and call analysis |
| Mistral: Saba Listed Feb 17, 2025 Official Mistral docs describe Saba as a fine-tuned 24B regional model for the Middle East and South Asia, and Mistral later marked it retired in favor of Mistral Small 4. Strong in its niche, but not a top-tier general-purpose model. | Mistral | 5.8/10 | $0.200 | $0.600 | 33k | Regional-language specialization, Arabic and South Asian language focus, Useful for culturally grounded responses, Officially documented lifecycle and retirement | Arabic and South Asia localization, Regional customer support, Localized summarization and Q&A, Domain adaptation for regional contexts |
| AionLabs: Aion-2.0 Listed Feb 23, 2026 Aion Labs positions Aion 2.0 as a DeepSeek V3.2 variant optimized for immersive roleplaying and storytelling, with special strength in tension and conflict. The available evidence is mainly descriptive, so general-purpose capability and coding strength remain only moderately supported. | AionLabs | 5.7/10 | $0.800 | $1.600 | 131k | Officially described as a DeepSeek V3.2 variant., Docs emphasize immersive roleplaying and storytelling quality., Particularly noted for introducing tension, crises, and conflict., Reasoning support is exposed in the Aion Labs API. | Immersive roleplaying and story generation., Writing scenes that need tension or conflict., General assistant use where creative prose matters more than technical depth. |
| OpenAI: GPT Audio Listed Jan 19, 2026 OpenAI documents GPT-Audio as a generally available audio model for audio-in/audio-out use. It appears optimized for voice interaction rather than frontier text reasoning or coding, so I rate it as competent but specialized. | OpenAI | 5.7/10 | $2.500 | $10.00 | 128k | Audio input and output, Good for voice-agent experiences, Supports function calling, Structured outputs available in the API | Speech-to-speech assistants, Voice agents, Audio transcription/response workflows, Natural-sounding spoken interactions |
| Mistral: Ministral 3 3B 2512 Listed Dec 2, 2025 The evidence points to a competent but compact edge model: useful and flexible, yet clearly below larger general-purpose models in depth. Good for routine multimodal and structured tasks, not for demanding reasoning-heavy work. | Mistral | 5.7/10 | $0.100 | $0.100 | 131k | Smallest Ministral 3 model with native vision support and 256k context., Supports structured outputs, function calling, document Q&A, and batching., Designed for edge deployment and efficiency., Best suited to lightweight multimodal tasks rather than heavy reasoning. | Tiny edge assistant with image input., Simple extraction and classification tasks., Long-context but low-cost document processing., Tool-calling for constrained environments. |
| Morph: Morph V3 Fast Listed Jul 7, 2025 Morph V3 Fast is optimized for speed and code application, not broad reasoning. The evidence supports strong utility in edit-apply loops, but general assistant performance should be considered narrow. | Morph | 5.7/10 | $0.800 | $1.200 | 82k | Fastest Morph apply model, Purpose-built for real-time code edits, High reported merge accuracy, Long context for large-file edits | Low-latency code patching, Interactive code-edit assistants, High-throughput apply steps in agents, Rapid code merge operations |
| Amazon: Nova Lite 1.0 Listed Dec 5, 2024 AWS documents Nova Lite as a low-cost multimodal model for text, image, and video tasks with a 300K context window, but not standout deep reasoning. | Amazon | 5.7/10 | $0.060 | $0.240 | 300k | AWS describes it as a low-cost multimodal text/image/video model., 300K context window with text output and enterprise deployment support., Built for fast document analysis and visual Q&A., Service-card documentation from AWS rather than marketing-only claims. | Document and media analysis., Fast multimodal extraction and Q&A., Cost-sensitive enterprise assistants., Workflow routing and summarization. |
| Relace: Relace Search Listed Dec 8, 2025 The public evidence suggests a specialized code-retrieval and agentic-search tool, not a broad general-purpose model. I can justify moderate coding/search utility, but there is insufficient independent evidence to score it confidently across general reasoning or instruction following. | Relace | 5.6/10 | $1.000 | $3.000 | 256k | Purpose-built for codebase exploration and retrieval., Uses parallel file-view and grep-style tools to surface relevant files., Documented as agentic search rather than generic LLM chat., Good fit for repository navigation and context gathering. | Finding relevant files in large codebases., Agentic repository search before code edits., Coding-assistant context gathering., Search-driven code review support. |
| AionLabs: Aion-3.0-Mini Listed Jul 7, 2026 Official Aion Labs docs describe this as a DeepSeek-based multi-model system optimized for narrative structure, tension, and darker themes, with controllable reasoning. Evidence is strong for storytelling focus but limited for general-purpose coding or broad benchmarked capability. | AionLabs | 5.5/10 | $0.700 | $1.400 | 131k | Built on the DeepSeek family with a collaborative multi-model generation process., Official docs say it is tuned for stronger narrative structure and more compelling tension/conflict., Reasoning is supported via the Aion Labs API and can be controlled with reasoning_effort., Available through an OpenAI-compatible API, which helps integration and tool use. | Roleplay and interactive storytelling., Creative writing with darker or higher-tension scenes., API-based assistant workflows where story quality matters more than broad coding strength. |
| Google: Nano Banana Pro (Gemini 3 Pro Image) Listed Jun 18, 2026 Google’s model card describes Gemini 3 Pro Image as the more advanced, reasoning-driven image model for studio-quality editing, grounded visuals, and accurate text rendering. It is the strongest Google image model in this set for complex visual tasks, but it remains specialized to image work rather than coding or open-ended agent use. | 5.5/10 | $2.000 | $12.00 | 131k | Professional-grade image editing and generation, Real-world grounding via Google Search, Strong text rendering for graphics and data visualizations, Supports search grounding and thinking | High-fidelity design and editorial graphics, Product mockups and ad creatives, Factual visualizations needing accurate labels and grounding | |
| TheDrummer: Cydonia 24B V4.1 Listed Sep 27, 2025 The evidence mostly shows this is a 24B Mistral-derived fine-tune with a creative-writing focus, but there are no strong public benchmarks here. I rate it as competent for routine text tasks, with limited basis to claim deeper reasoning or coding strength. | TheDrummer | 5.5/10 | $0.300 | $0.500 | 131k | built on Mistral Small 3.2 24B, aimed at creative writing and roleplay-style output, appears to emphasize recall and prompt adherence, likely usable for everyday drafting and formatting | creative writing and persona-style chat, routine instruction-following and structured text, light coding assistance where creativity matters more than rigor |
| Relace: Relace Apply 3 Listed Sep 26, 2025 Relace Apply 3 is clearly specialized for code patch application, and the official materials emphasize merge accuracy, speed, and long context. That supports high coding utility for diff-merging, but not broad reasoning or general instruction-following beyond its niche. | Relace | 5.5/10 | $0.850 | $1.250 | 256k | specialized for applying code diffs/merges, designed for high-speed patching workflows, focused on merge accuracy across code-edit formats, large context for handling big files and edits | automated code patch application, agentic coding pipelines and edit merging, high-throughput repo maintenance workflows |
| Sao10K: Llama 3 8B Lunaris Listed Aug 13, 2024 Public evidence is mostly catalog-style, but the underlying Llama-3.1-8B base and nearby Lunaris model pages suggest a competent small generalist with creative leaning. I would not assume frontier-level reasoning or robustness without stronger evals. | Sao10K | 5.5/10 | $0.040 | $0.050 | 8k | Community model with explicit chat/roleplay orientation., Based on Llama 3 8B, so likely solid on routine generative tasks., Public listings indicate 8K context and chat focus. | Roleplay and creative chat, General-purpose lightweight assistant, Prompt-formatting and structured generation |
| Moondream: Moondream 3 Preview Listed Oct 15, 2026 Moondream 3 Preview is a capable small VLM with strong grounding, OCR, and document/UI understanding, backed by broad public benchmarks and native structured skills. | Moondream | 5.4/10 | $0.300 | $2.500 | 32k | Strong visual reasoning and grounded VQA for image understanding tasks., Native detection, pointing, captioning, OCR, and segmentation with structured outputs., 32K context and reasoning mode help with multi-step image analysis., Public benchmarks include RefCOCO, ChartQA, DocVQA, TextVQA, POPE, and ScreenSpot. | Visual question answering over real-world images and documents., Object detection, pointing, and segmentation in production vision workflows., OCR, chart/document understanding, and UI/screen analysis., Fine-tuning a compact vision-language model for a specialized visual task. |
| Reka Edge Listed Mar 20, 2026 Reka Edge is clearly positioned as an efficient 7B multimodal vision-language model for edge deployment and visual grounding. Strong for perception-heavy tasks, but its smaller size and specialization suggest more limited general reasoning and coding than frontier text models. | rekaai | 5.4/10 | $0.100 | $0.100 | 16k | multimodal image/video understanding, edge-friendly efficiency, tool use and structured outputs, good physical-grounding orientation | vision-language assistant tasks, video/image analysis, edge deployment, object grounding and visual QA |
| Google: Nano Banana 2.1 Listed Oct 6, 2026 Official Google docs describe it as the latest high-efficiency image generation/editing model with improved visual quality, typography, and multi-turn consistency. It looks strong for visual workflows, but it is not a general-purpose text/code model, so overall reasoning and coding are limited. | 5.3/10 | $1.500 | $7.500 | 66k | Strong image generation and conversational editing, Multi-turn consistency and prompt adherence, High-resolution output support and text rendering, Multi-image fusion / reference-image workflows | Image editing and retouching, Ad creative and product mockups, Storyboarding and visual iteration, Infographics and text-heavy visual layouts | |
| AionLabs: Aion 3.5 Mini Listed Sep 23, 2026 The evidence points to a storytelling-focused system built on GLM, not a broadly verified general-purpose model. Without independent benchmarks or a model card, I can only rate it as competent for narrative tasks and uncertain outside that niche. | AionLabs | 5.3/10 AI review detailsLow confidence · Sep 23, 2026 Coding: 4.8/10 Reasoning: 5.4/10 Reliability: 5.1/10 Instructions: 5.7/10 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $0.700 | $1.400 | 262k | Designed for roleplaying/storytelling, Collaborative multi-model generation process, Tool calling and JSON structured outputs supported, Large 262K-token context window | Creative writing and roleplay, Narrative generation with long context, Interactive fiction or character-driven dialogue, Tool-augmented text workflows |
| AionLabs: Aion 3.5 Listed Sep 23, 2026 Evidence supports a specialized storytelling system with long context and reasoning support, but there are no substantive public benchmark results here for coding or broad general-purpose strength. I’d treat it as competent for narrative tasks, with notable uncertainty outside that specialty. | AionLabs | 5.3/10 AI review detailsMedium confidence · Sep 23, 2026 Coding: 3.5/10 Reasoning: 6.5/10 Reliability: 5/10 Instructions: 6/10 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $3.000 | $6.000 | 262k | Very long 256K context window for extended sessions and documents., Built as a multi-model GLM-family system with collaborative generation., Documented focus on narrative structure, tension/conflict, and mature/darker themes., Reasoning mode and 32K max output are explicitly documented. | Roleplaying and interactive fiction., Long-form storytelling with persistent context., Creative writing with stronger scene structure and conflict., General text tasks where narrative quality matters more than technical precision. |
| Google: Nano Banana 2 (Gemini 3.1 Flash Image Preview) Listed Feb 26, 2026 Google’s official docs identify Gemini 3.1 Flash Image (Nano Banana 2) as an image understanding/generation model with configurable thinking and search grounding. Because this is a preview image model rather than a general-purpose LLM, coding value is limited and overall breadth is narrower. | 5.3/10 | $0.500 | $3.000 | 66k | Strong image generation and editing focus, Supports grounded image workflows, Configurable thinking on Google’s platform, Designed for high-fidelity visual creation | Image generation, Image editing, Reference-image composition, Visually grounded creative workflows | |
| Meta: Llama 3.2 3B Instruct Listed Sep 25, 2024 Meta’s official card places the 3B model in the compact-but-capable tier, with substantially better benchmark results than the 1B model but still clear gaps versus larger Llama 3.2 variants. I’d rate it as competent for routine tasks, not high-end reasoning. | Meta | 5.3/10 | $0.050 | $0.330 | 131k | Better balanced small-model performance than the 1B variant., Strong for multilingual dialogue and summarization., Works well in constrained environments with 128k context in the family card., Reasonable instruction-following for its size. | Small assistant apps, Summarization and rewriting, Multilingual Q&A, On-device or edge inference |
| Tencent: Hy-MT2-30B-A3B Listed Aug 20, 2026 Evidence supports a strong specialist translation model with robust constraint-following, but not a general-purpose reasoner. Public sources do not provide enough independent, exact benchmark data to justify stronger claims about coding or broad reasoning. | Tencent | 5.2/10 | $0.074 | $0.295 | 8k | Strong multilingual translation focus across 33 languages, Designed to preserve structure and follow detailed translation constraints, MoE 30B-A3B variant aimed at higher translation quality than smaller Hy-MT2 models, Official examples cover structured data, delimiters, terminology, and style control | High-volume multilingual translation, Structured-content translation where keys/placeholders must be preserved, Terminology/glossary-constrained localization, Instruction-heavy translation workflows and subtitle-style adaptation |
| NVIDIA: Nemotron 3.5 Content Safety Listed Jun 4, 2026 NVIDIA positions this as a 4B Gemma-3-4B-it-based multimodal safety moderator with custom-policy reasoning. It looks strong for guardrails, but remains specialized rather than general-purpose. | NVIDIA | 5.2/10 | $0.200 | $0.200 | 131k | Multimodal safety moderation over text and image inputs, including assistant responses., Custom-policy mode with concise reasoning traces for user-defined safety rules., Strong specialization for jailbreak/content-classification workflows., Backed by NVIDIA’s own evaluations on guardrail and false-positive benchmarks. | LLM/VLM input and output moderation., Enterprise policy enforcement with custom taxonomies., Jailbreak and prompt-injection screening., Safety gating in agent or chatbot pipelines. |
| NVIDIA: Nemotron 3.5 Content Safety (free) Listed Jun 4, 2026 This appears to be the same Nemotron 3.5 Content Safety model exposed through a free access tier. I score it the same, while noting the free listing adds no broader-capability evidence. | NVIDIA | 5.2/10 | $0.000 | $0.000 | 128k | Multimodal safety moderation over text and image inputs, including assistant responses., Custom-policy mode with concise reasoning traces for user-defined safety rules., Strong specialization for jailbreak/content-classification workflows., Backed by NVIDIA’s own evaluations on guardrail and false-positive benchmarks. | LLM/VLM input and output moderation., Enterprise policy enforcement with custom taxonomies., Jailbreak and prompt-injection screening., Safety gating in agent or chatbot pipelines. |
| OpenAI: GPT Audio Mini Listed Jan 19, 2026 OpenAI describes GPT-Audio Mini as a cost-efficient audio model, but also notes limitations versus fuller-featured models. I expect solid voice-task handling, while general reasoning and coding remain secondary to the audio modality. | OpenAI | 5.2/10 | $0.600 | $2.400 | 128k | Cost-efficient audio model, Audio input and output, Function calling support, Useful for lightweight voice apps | Budget voice agents, Simple speech workflows, Audio-enabled assistants, High-volume spoken interactions |
| TheDrummer: Skyfall 36B V2 Listed Mar 10, 2025 Evidence suggests this is a Mistral Small 2501-derived fine-tune aimed at creativity/RP, not a frontier generalist. Exact independent benchmark coverage is sparse, so capability estimates are cautious and mostly inferential. | TheDrummer | 5.2/10 | $0.550 | $0.800 | 33k | Creative writing and role-play tuning, Coherent long-form storytelling, Mid-size text model for general chat, Can use structured output in hosted wrappers | Fiction drafting and character dialogue, Role-play sessions, Style-heavy rewriting, General assistant tasks where creativity matters more than technical depth |
| OpenAI: GPT-5.4 Image 2 Listed Apr 21, 2026 I could not find a clean official page for the exact GPT-5.4-image-2 identifier, so this is inferred from OpenAI’s GPT Image 2 / ChatGPT Images 2.0 materials. It looks strong for image creation and editing, but not like a general coding model; exact capability for this alias remains uncertain. | OpenAI | 5.1/10 | $8.000 | $15.00 | 272k | Image generation/editing focus, Can integrate reasoning and web-grounded workflows, Good for dense-text and instruction-heavy image work, Best suited to multimodal creative tasks | Image generation, Image editing, Presentation and marketing visuals, Reasoned multimodal image workflows |
| OpenAI: gpt-oss-safeguard-20b Listed Oct 29, 2025 OpenAI’s technical report says gpt-oss-safeguard-20b is a fine-tuned open-weight reasoning model for labeling content against a supplied policy, not a general assistant. Its reliability is strongest in safety-classification workflows. | OpenAI | 5.1/10 | $0.075 | $0.300 | 131k | Purpose-built for safety classification, Supports custom policies and structured outputs, Open-weight and customizable | Content policy classification, LLM output filtering, Trust-and-safety workflows |
| ReMM SLERP 13B Listed Jul 22, 2023 This is a recreation/merge model with moderate public evidence: its HF card shows Open LLM Leaderboard results around mid-pack for 13B open models, while the repo framing suggests a MythoMax-like recreation. Stronger tasks remain uncertain. | undi95 | 5.1/10 | $0.350 | $0.650 | 6k | Open LLM Leaderboard results are available., Can produce structured outputs reasonably well., Based on a proven Llama-2-era 13B lineage. | Light assistant/chat use, Structured text generation, Offline experimentation with a 13B model |
| Google: Nano Banana 2 (Gemini 3.1 Flash Image) Listed Jun 18, 2026 Google’s model card and release notes show Gemini 3.1 Flash Image as a GA native visual model with grounding, higher resolutions, and improved consistency. It appears stronger than the Lite variant for demanding visual work, but it is still an image model rather than a coding or general agent model. | 5.0/10 | $0.500 | $3.000 | 131k | Higher-quality image generation and conversational editing, Supports search grounding for real-world context, Supports 2K and 4K outputs, Better text rendering and aspect-ratio adherence | Professional-grade image editing and generation, Marketing visuals and product mockups, Grounded visual content with web-backed context | |
| OpenAI: GPT-3.5 Turbo 16k Listed Aug 28, 2023 OpenAI’s release notes describe this as a 2023-era 16k-context upgrade, but it is now a deprecated legacy snapshot. It is somewhat better than the base GPT-3.5 chat snapshots on context, yet still well behind modern models in reasoning and reliability. | OpenAI | 4.8/10 | $3.000 | $4.000 | 16k | Longer legacy context window than early GPT-3.5 chat models., Works for basic chat and non-chat tasks., Supported function calling in the 2023-era family update., Useful only for legacy long-context compatibility now. | Legacy long-context chat, Older app migrations, Basic summarization over mid-sized inputs, Function-calling on old systems |
| MythoMax 13B Listed Jul 2, 2023 Evidence points to an older Llama 2 13B fine-tune aimed at chat and roleplay, not a modern general-purpose model. Public evidence for coding and reasoning is limited, so capability looks competent but clearly dated. | gryphe | 4.8/10 | $0.080 | $0.110 | 8k | Good roleplay and prose style for its class, Can follow straightforward chat instructions, Popular fine-tune with long community usage history | Creative writing and roleplay, Lightweight chat on older hardware, Basic summarization or drafting |
| Perceptron: Perceptron Mk1 Listed May 12, 2026 Perceptron Mk1 is clearly specialized for video and embodied reasoning, not broad software or text work. The public evidence supports that specialization well, but it does not justify strong general-purpose coding or reasoning scores. | Perceptron | 4.7/10 | $0.150 | $1.500 | 33k | Strong video and embodied reasoning focus, Native structured spatial outputs like point/box/track/clip, Useful for OCR, document extraction, and visual search, Official announcement gives concrete modality and context details | Video understanding, Robotics and embodied perception, Visual search and inspection, Document/image extraction |
| Google: Nano Banana (Gemini 2.5 Flash Image) Listed Oct 7, 2025 This is a specialized image model, so I score coding very low because that capability is not the point of the model. For image-centric work it looks strong and current, but it is not a general-purpose reasoning model. | 4.7/10 | $0.300 | $2.500 | 33k | State-of-the-art image generation and editing focus, Native multimodal understanding for creative workflows, Low-latency, high-volume creation orientation, Officially GA and documented by Google | Image generation at scale, Conversational image editing, Creative iteration workflows, Product mockups and visual ideation | |
| ByteDance: UI-TARS 7B Listed Jul 22, 2025 Specialized GUI agent with vision-language input and reported strength in computer-use tasks. Useful for interface automation, though public evidence is thinner than for frontier general-purpose models. | ByteDance | 4.7/10 | $0.100 | $0.200 | 128k | GUI-focused multimodal interaction, Desktop/web/mobile/game environment handling, Vision-language input, Structured agent-style outputs | Automating repetitive GUI workflows, Browser or desktop task execution, Screenshot-based UI inspection, Research prototypes for computer-use agents |
| OpenAI: GPT-3.5 Turbo (older v0613) Listed Jan 25, 2024 OpenAI’s own posts show this snapshot was a 2023-era capability upgrade, but it is now deprecated and materially behind current models. It remains a legacy baseline with moderate instruction following and limited reasoning/coding by modern standards. | OpenAI | 4.7/10 | $1.000 | $2.000 | 4k | Introduced function calling and better steerability for its time., Still usable for simpler chat and transformation tasks., Legacy chat-model compatibility., Cheap and light compared with larger models when it was current. | Legacy app compatibility, Simple chatbots, Basic extraction/formatting, Function-calling prototypes on old stacks |
| Inference.net: Schematron V2 Small Listed Sep 12, 2026 Schematron V2 Small is also a 3B extraction-specialist model, described as the family’s highest-quality option and aimed at complex schemas and long pages. That supports better extraction fidelity than Turbo, but still not strong general-purpose reasoning or coding. | Inference.net | 4.5/10 | $0.050 | $0.230 | 128k | Highest extraction quality in the Schematron family, Specialized for messy HTML to clean JSON conversion, Schema-constrained decoding with valid JSON by construction, Long-context support for complex pages | Complex HTML extraction, Typed data generation for long web pages, Schema-heavy ETL workflows, High-accuracy structured output tasks |
| Inference.net: Schematron V2 Turbo Listed Sep 12, 2026 Schematron V2 Turbo is a 3B specialized extraction model optimized for speed and cost, not a general reasoning or coding model. The evidence supports strong instruction adherence for schema-constrained extraction, but not broad capability outside that narrow task. | Inference.net | 4.4/10 | $0.030 | $0.150 | 128k | Very strong HTML-to-JSON extraction focus, High-throughput design for scale, Schema-constrained outputs reduce formatting errors, Simple OpenAI-compatible structured-output workflow | Large-scale web extraction pipelines, Product/feed normalization, Schema-first data extraction from messy pages, High-volume JSON generation from HTML |
| TheDrummer: UnslopNemo 12B Listed Nov 8, 2024 Evidence points to a niche role-play and creative-writing fine-tune rather than a broadly capable general model. Confidence is only moderate because public artifacts are thin and lack rigorous benchmark evidence. | TheDrummer | 4.3/10 | $0.400 | $0.400 | 1M | Fine-tuned for role-play and adventure writing., Can be expressive in conversational creative scenarios., Community artifacts suggest a 12B llama-based model., Useful for structured chat-style outputs when prompted carefully. | Roleplay and character dialogue, Interactive fiction or adventure writing, Creative brainstorming, Style-focused chat experiences |
| Tencent: Hy-MT2-7B Listed Aug 19, 2026 Hy-MT2-7B should be stronger than the 1.8B sibling on translation quality and robustness, but it remains a specialist translation system. The available evidence supports high translation capability, not broad coding or agentic performance. | Tencent | 4.2/10 | $0.074 | $0.295 | 8k | Translation-specialized model with stronger capacity than 1.8B, Strong multilingual instruction following, Supports 33 languages, Suitable for production translation workflows | high-quality translation, enterprise localization, multilingual instruction-following tests, subtitle and document translation |
| Google: Nano Banana 2 Lite (Gemini 3.1 Flash Lite Image) Listed Jun 30, 2026 Google’s model card positions Gemini 3.1 Flash Lite Image as an efficiency-focused image model with sub-2-second latency, 1K output, and multimodal inputs. It is strong for fast visual workflows, but the docs do not indicate coding or broad general-purpose agent capabilities. | 4.2/10 | $0.250 | $1.500 | 66k | Very low-latency image generation and editing, Supports text, image, video, and PDF inputs, Native image output with 1K resolution, Suitable for high-volume interactive pipelines | Rapid image generation/editing at scale, Interactive consumer image tools, Fast visual exploration and lightweight multimodal workflows | |
| OpenAI: GPT-3.5 Turbo Instruct Listed Sep 28, 2023 OpenAI describes this as an older completions-only model with similar capabilities to GPT-3-era models. It is clearly a legacy model today, useful mainly for backward compatibility rather than strong general capability. | OpenAI | 4.2/10 | $1.500 | $2.000 | 4k | Legacy completions-only interface., Suitable for simple prompt-completion workflows., Stable snapshot behavior for old applications., Lower-complexity text generation and transformation. | Legacy completions migration support, Text generation in old pipelines, Simple rewriting/classification, Compatibility with older integrations |
| ScaleDown: Summarize Listed Oct 13, 2026 Evidence only supports that this is a ScaleDown prompt-compression/summarization offering; I found no public benchmark for the exact model, so capability is inferred from the task focus and repo/docs. Strongest fit is summarization, while coding and broader reasoning remain unverified and likely narrow. | ScaleDown | 4.1/10 | $0.040 | $0.000 | 1M | Purpose-built for abstractive summarization tasks., Likely effective for long-document condensation and chat-log summarization., Designed to preserve semantic meaning while compressing text., Fits workflows where concise, fluent rewrites are needed. | Summarizing long documents into shorter prose., Condensing meeting notes or conversation histories., Generating executive summaries from verbose source text. |
| ScaleDown: Extract Listed Oct 13, 2026 Evidence supports an extraction-specialized ScaleDown task model, but I found no public evaluation for this exact version, so the score is a cautious inference rather than a measured result. It should be competent for structured extraction, but general reasoning, coding, and reliability outside extraction are not established. | ScaleDown | 4.1/10 | $0.040 | $0.000 | 1M | Built for description-driven entity/field extraction., Can adapt to caller-specified schemas instead of fixed entity labels., Well suited to structured output from unstructured text., Likely useful for document triage and information harvesting. | Extracting custom fields from contracts, forms, or reports., Turning unstructured text into JSON records., Pulling targeted entities or attributes from long documents. |
| Tencent: Hy-MT2-1.8B Listed Aug 20, 2026 Hy-MT2-1.8B is an excellent translation-specialist model, but the official paper and repo scope it to multilingual translation rather than broad general intelligence. I would not transfer its translation strength to coding or general reasoning. | Tencent | 3.9/10 | $0.044 | $0.177 | 8k | Specialized multilingual translation model, Strong translation instruction-following focus, Compact on-device footprint, Supports many language pairs | machine translation, localization pipelines, subtitle translation, domain-specific bilingual workflows |
| AionLabs: Aion-RP 1.0 (8B) Listed Feb 4, 2025 The model card explicitly says this is an uncensored roleplay and creative-writing fine-tune, which supports a narrow specialization rather than broad capability. There is not enough solid public benchmark evidence to rate it highly beyond that niche. | AionLabs | 3.9/10 AI review detailsMedium confidence · Oct 17, 2026 Coding: 3.2/10 Reasoning: 3.8/10 Reliability: 3.9/10 Instructions: 5/10 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $0.800 | $1.600 | 33k | Clearly tuned for roleplay and creative writing., Long context support is advertised in the model card., May sustain multi-turn character consistency better than a base model., Likely useful for uncensored conversational RP use. | Roleplay character chat, Creative writing assistance, Dialogue generation, Light multi-turn immersive scenarios |
| Mancer: Weaver (alpha) Listed Aug 2, 2023 Evidence is thin: the public listing describes it as a Claude-style verbosity recreation aimed at roleplay/narrative, but provides no substantive benchmark data. Capability is therefore inferred conservatively, with likely weaknesses in coherence and memory. | Mancer | 3.5/10 | $0.400 | $0.750 | 8k | Produces verbose, narrative-style generations., May fit roleplay or creative prose tasks., Easy to prompt for a specific voice or style. | Roleplay dialogue, Story or character writing, Style-heavy brainstorming |
| Meta: Llama 3.2 1B Instruct Listed Sep 25, 2024 Meta’s model card shows this 1B model is intended for lightweight multilingual assistant use and reports much lower benchmark performance than the 3B and 8B family members. It is best viewed as a compact utility model, not a strong general reasoner. | Meta | 3.3/10 | $0.027 | $0.201 | 60k | Very small footprint and efficient deployment., Designed for multilingual dialogue, retrieval, and summarization., Useful on-device / constrained-resource scenarios., Can handle basic assistant tasks despite size. | On-device assistants, Simple summarization, Prompt rewriting, Low-latency embedded workflows |
| Google: Lyria 3 Pro Preview Listed Mar 30, 2026 Lyria 3 Pro Preview is a specialized music-generation model, not a general LLM. It is well documented for song composition and structural control, but it has no meaningful coding or broad reasoning role outside audio generation. | 2.7/10 | $0.000 | $0.000 | 1M | full-song music generation, text and image-to-music input, vocal/instrumental control, structure controls for songs | full-length song generation, music ideation, soundtrack prototyping, lyrics-driven music creation | |
| Meta: Llama Guard 4 12B Listed Apr 30, 2025 Strong evidence supports Llama Guard 4 as a specialized multimodal moderation model, not a general assistant. Best for safety filtering; it is not suited to open-ended reasoning or coding. | Meta | 2.7/10 AI review detailsHigh confidence · Oct 17, 2026 Coding: 0.5/10 Reasoning: 2/10 Reliability: 8/10 Instructions: 4/10 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $0.180 | $0.180 | 164k | Multimodal safety classification on text and images, Supports input and output moderation, Covers MLCommons hazard taxonomy, Designed for multilingual moderation workflows | Content moderation for chat and image inputs, Pre-filtering prompts before generation, Post-filtering model outputs, Safety gating in assistant pipelines |
| Google: Lyria 3 Clip Preview Listed Mar 30, 2026 Lyria 3 Clip Preview is a specialized short-form music model, optimized for 30-second clips rather than general language tasks. Evidence supports strong audio-generation control, but not general reasoning or coding capability. | 2.6/10 | $0.000 | $0.000 | 1M | 30-second music clips, text and image-to-music input, vocal and instrumental modes, audio watermarking and content credentials | short music clips, loops and previews, creative audio experiments, rapid music ideation | |
| Elorian: Glimpse Listed Oct 17, 2026 Public evidence I found is limited to the OpenRouter catalog entry, which describes image/video reasoning and a 262K context window. I could not find independent benchmark results or a model card, so I can’t score capability confidently. | Elorian | Limited evidence AI review detailsLow confidence · Oct 17, 2026 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $0.600 | $3.000 | 262k | Native image/video input, Claims support for tool calling and structured output, Very large 262K context window, Positioned for visual search, captioning, and dense counting | Visual search on images and video, Video captioning and event tracking, Counting or inspection in dense scenes, Agentic multimodal workflows |
| Moondream: Moondream 3.1 Listed Oct 15, 2026 No separate assessment was provided for this model in the draft. | Moondream | Limited evidence AI review detailsLow confidence · Oct 15, 2026 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $0.300 | $2.500 | 32k | image input, reasoning controls | — |
| Meta: Muse Spark Open Listed Oct 15, 2026 Meta’s launch post says Muse Spark Open is a new open-weight frontier model built for coding and agentic work. I couldn’t find enough independent, exact-model benchmark evidence to score it confidently yet, so this is provisional and may change as third-party evaluations appear. | Meta | Limited evidence | $1.250 | $4.250 | 1M | Open-weight release, Natively multimodal, Large context window, Designed for coding and agentic workflows | Local/self-hosted experimentation, Coding agents, Long-context multimodal tasks, Research into open-weight deployment |
| Anthropic: Claude Haiku 5.5 Listed Oct 7, 2026 Anthropic publicly mentioned Claude Haiku 5.5 would arrive later, but I could not find a system card, benchmark report, or API documentation for the exact model. Without primary evidence for the exact release, I can’t responsibly assign capability scores. | Anthropic | Limited evidence | $0.100 | $0.500 | 1M | Likely intended for fast, high-volume workflows based on Anthropic’s family positioning, but this is, not verified for the exact model version., May be suitable for lightweight agent/tool-use tasks if Anthropic ships it as described, but no, public model card or benchmark evidence for the exact release was found. | High-volume summarization, Browser/tool-using agents, Cost-sensitive routine assistant workflows |
| Anthropic: Claude Sonnet 5.5 Listed Sep 28, 2026 No assessment for Sonnet 5.5 is present in the supplied draft, so no scores can be reliably extracted. | Anthropic | Limited evidence | $2.000 | $10.00 | 1M | image input, tool calling, reasoning controls, structured output | — |
| Anthropic: Claude Sonnet 5.5 (batch) Listed Sep 28, 2026 No assessment for Sonnet 5.5 is present in the supplied draft, so no scores can be reliably extracted. | Anthropic | Limited evidence | $1.000 | $5.000 | 1M | image input, tool calling, reasoning controls, structured output | — |
| Qwen: Qwen3.8 Max Prime Listed Sep 23, 2026 I couldn’t find a primary-source model card, benchmark, or vendor doc for the exact `qwen/qwen3.8-max-prime` SKU. Because the evidence only supports the broader Qwen3.8-Max family, I can’t responsibly score this exact model’s capability. | Qwen | Limited evidence | $4.000 | $12.00 | 1M | Possible relation to Qwen3.8-Max’s multimodal, 1M-context flagship line, but the exact SKU is not公开/, OpenRouter catalog suggests higher-throughput positioning, though this is unverified and not a first, No public model card, benchmark table, or vendor docs were found for the exact `prime` variant., Exact capabilities may differ materially from Qwen3.8-Max and cannot be safely inferred. | Only if your own testing confirms behavior matches the intended Qwen3.8-Max Prime SKU., Internal experimentation where you can benchmark the exact deployed endpoint yourself. |
| OpenAI: GPT-6 Luna Listed Sep 22, 2026 No primary-source evidence found for this exact model ID. OpenAI’s official API docs list GPT-6 Astra and GPT-5.6 family models, not GPT-6 Luna, so capability can’t be assessed reliably. | OpenAI | Limited evidence | $0.100 | $0.500 | 1.1M | image input, tool calling, reasoning controls, structured output | — |
| OpenAI: GPT-6 Luna (batch) Listed Sep 22, 2026 No primary-source evidence found for this exact batch variant. OpenAI’s official API docs list GPT-6 Astra and GPT-5.6 family models, not GPT-6 Luna batch, so capability can’t be assessed reliably. | OpenAI | Limited evidence | $0.050 | $0.250 | 1.1M | image input, tool calling, reasoning controls, structured output | — |
| OpenAI: GPT-6 Sol Listed Sep 22, 2026 No primary-source evidence found for this exact model ID. OpenAI’s official docs currently list GPT-6 Astra and GPT-5.6 Sol/Terra/Luna, but not GPT-6 Sol, so capability can’t be assessed reliably. | OpenAI | Limited evidence | $2.000 | $10.00 | 1.1M | image input, tool calling, reasoning controls, structured output | — |
| OpenAI: GPT-6 Sol (batch) Listed Sep 22, 2026 No primary-source evidence found for this exact batch variant. OpenAI’s official docs currently list GPT-6 Astra and GPT-5.6 Sol/Terra/Luna, but not GPT-6 Sol batch, so capability can’t be assessed reliably. | OpenAI | Limited evidence | $1.000 | $5.000 | 1.1M | image input, tool calling, reasoning controls, structured output | — |
| Xiaomi: MiMo-V2.6-Pro-UltraSpeed Listed Sep 21, 2026 I could not verify an official Xiaomi MiMo-V2.6-Pro-UltraSpeed release or benchmark evidence for this exact model ID. Xiaomi’s public docs instead cover MiMo-V2-Pro, MiMo-V2-Flash, and MiMo-V2.5 models, so any capability rating here would be speculative. ([mimo.mi.com](https://mimo.mi.com/docs/en-US/updates/model?utm_source=openai)) | Xiaomi | Limited evidence | $4.350 | $8.700 | 1M | Catalog suggests multimodal and tool-use oriented capabilities, but this is unverified for the exact, Exact Xiaomi sources found for 2026-09-21 discuss MiMo-V2-Pro, MiMo-V2-Flash, and MiMo-V2.5 variants, No official primary source located for a MiMo-V2.6-Pro-UltraSpeed release or benchmark set, UltraSpeed naming appears only in the supplied catalog and earlier V2.5 UltraSpeed references, not v | Not enough evidence to assess the exact model reliably, If the catalog entry is an alias, it would need direct vendor confirmation before use in production |
| OpenAI: GPT-6 Astra Listed Sep 4, 2026 Assessment is missing in the draft for GPT-6 Astra, so no score is provided here. | OpenAI | Limited evidence | $10.00 | $50.00 | 1.1M | image input, tool calling, reasoning controls, structured output | — |
| OpenAI: GPT-6 Astra (batch) Listed Sep 4, 2026 Assessment is missing in the draft for GPT-6 Astra, so no score is provided here. | OpenAI | Limited evidence | $5.000 | $25.00 | 1.1M | image input, tool calling, reasoning controls, structured output | — |
| OpenAI: GPT-6 Astra Pro Listed Sep 4, 2026 Assessment missing in the supplied draft. | OpenAI | Limited evidence AI review detailsLow confidence · Oct 6, 2026 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $10.00 | $50.00 | 1.1M | image input, tool calling, reasoning controls, structured output | — |
| OpenAI: GPT-6 Astra Pro (batch) Listed Sep 4, 2026 Assessment missing in the supplied draft. | OpenAI | Limited evidence AI review detailsLow confidence · Oct 6, 2026 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $5.000 | $25.00 | 1.1M | image input, tool calling, reasoning controls, structured output | — |
| Meta: Muse Spark 1.3 Listed Sep 2, 2026 No usable assessment was supplied for this model in the draft, so I’m leaving scores unset. | Meta | Limited evidence | $1.250 | $4.250 | 1M | image input, tool calling, reasoning controls, structured output | — |
| Anthropic: Claude Fable 5.1 Listed Sep 1, 2026 Assessment is missing in the draft for Claude Fable 5.1, so no score is provided here. | Anthropic | Limited evidence | $10.00 | $50.00 | 1M | image input, tool calling, reasoning controls, structured output | — |
| Anthropic: Claude Fable 5.1 (batch) Listed Sep 1, 2026 Assessment is missing in the draft for Claude Fable 5.1, so no score is provided here. | Anthropic | Limited evidence | $5.000 | $25.00 | 1M | image input, tool calling, reasoning controls, structured output | — |
| Meta: Muse Spark 1.2 Contributor Listed Aug 21, 2026 No usable assessment was supplied for this model in the draft, so I’m leaving scores unset. | Meta | Limited evidence | $0.100 | $0.200 | 1M | image input, tool calling, reasoning controls, structured output | — |
| Google: Gemini 3.7 Flash Listed Aug 13, 2026 No assessment was provided in the draft for this model. | Limited evidence AI review detailsLow confidence · Oct 17, 2026 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $0.750 | $3.750 | 1M | image input, audio input, tool calling, reasoning controls, structured output | — | |
| Google: Gemini 3.7 Flash (batch) Listed Aug 13, 2026 No assessment was provided in the draft for this model. | Limited evidence AI review detailsLow confidence · Oct 17, 2026 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $0.375 | $1.875 | 1M | image input, audio input, tool calling, reasoning controls, structured output | — | |
| Meta: Muse Spark 1.2 Listed Aug 5, 2026 No usable assessment was supplied for this model in the draft, so I’m leaving scores unset. | Meta | Limited evidence | $1.250 | $4.250 | 1M | image input, tool calling, reasoning controls, structured output | — |
| Qwen: Qwen3.7 Flash Listed Jul 27, 2026 I could not find a primary-source match for the exact model ID Qwen3.7 Flash. The closest official Qwen evidence is for Qwen3.7-Max/Plus and other Flash families, so capability for this exact ID is too uncertain to score. | Qwen | Limited evidence | $0.030 | $0.130 | 1M | image input, tool calling, reasoning controls | — |
| Anthropic: Claude Opus 5 Listed Jul 24, 2026 Assessment is missing in the draft for Claude Opus 5, so no score is provided here. | Anthropic | Limited evidence | $5.000 | $25.00 | 1M | image input, tool calling, reasoning controls, structured output | — |
| Anthropic: Claude Opus 5 (batch) Listed Jul 24, 2026 Assessment is missing in the draft for Claude Opus 5, so no score is provided here. | Anthropic | Limited evidence | $2.500 | $12.50 | 1M | image input, tool calling, reasoning controls, structured output | — |
| Poolside: Laguna S 2.1 Listed Jul 21, 2026 No assessment was provided in the draft for this model. | Poolside | Limited evidence AI review detailsLow confidence · Oct 17, 2026 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $0.090 | $0.180 | 1M | tool calling, reasoning controls | — |
| Poolside: Laguna S 2.1 (free) Listed Jul 21, 2026 No assessment was provided in the draft for this model. | Poolside | Limited evidence AI review detailsLow confidence · Oct 17, 2026 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $0.000 | $0.000 | 262k | tool calling, reasoning controls | — |
| Google: Gemini 3.6 Flash Listed Jul 21, 2026 No assessment was provided in the draft for this model. | Limited evidence AI review detailsLow confidence · Oct 17, 2026 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $0.750 | $3.750 | 1M | image input, audio input, tool calling, reasoning controls, structured output | — | |
| Google: Gemini 3.6 Flash (batch) Listed Jul 21, 2026 No assessment was provided in the draft for this model. | Limited evidence AI review detailsLow confidence · Oct 17, 2026 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $0.375 | $1.875 | 1M | image input, audio input, tool calling, reasoning controls, structured output | — | |
| Meta: Muse Spark 1.1 Listed Jul 16, 2026 No usable assessment was supplied for this model in the draft, so I’m leaving scores unset. | Meta | Limited evidence | $1.250 | $4.250 | 1M | image input, tool calling, reasoning controls, structured output | — |
| OpenAI: GPT-5.6 Sol Pro Listed Jul 9, 2026 Assessment missing in the supplied draft. | OpenAI | Limited evidence AI review detailsLow confidence · Oct 6, 2026 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $4.000 | $20.00 | 1.1M | image input, tool calling, reasoning controls, structured output | — |
| OpenAI: GPT-5.6 Sol Pro (batch) Listed Jul 9, 2026 Assessment missing in the supplied draft. | OpenAI | Limited evidence AI review detailsLow confidence · Oct 6, 2026 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $2.000 | $10.00 | 1.1M | image input, tool calling, reasoning controls, structured output | — |
| OpenAI: GPT-5.6 Sol Listed Jul 9, 2026 Assessment is missing in the draft for GPT-5.6 Sol, so no score is provided here. | OpenAI | Limited evidence | $4.000 | $20.00 | 1.1M | image input, tool calling, reasoning controls, structured output | — |
| OpenAI: GPT-5.6 Sol (batch) Listed Jul 9, 2026 Assessment is missing in the draft for GPT-5.6 Sol, so no score is provided here. | OpenAI | Limited evidence | $2.000 | $10.00 | 1.1M | image input, tool calling, reasoning controls, structured output | — |
| OpenAI: GPT Chat Latest Listed May 5, 2026 This is an alias, not a fixed model version, so there is no stable exact capability basis to score. OpenAI documents that it always resolves to the latest Instant model used in ChatGPT, which is too mutable for a version-specific review. | OpenAI | Limited evidence AI review detailsLow confidence · Oct 17, 2026 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $5.000 | $30.00 | 400k | Alias is documented by OpenAI as pointing to the latest Instant chat model used in ChatGPT | Cannot assess as an exact version, Alias may change over time |
| Arcee AI: Trinity Large Thinking Listed Apr 1, 2026 No assessment was provided in the draft for this model. | Arcee AI | Limited evidence AI review detailsLow confidence · Oct 17, 2026 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $0.250 | $0.800 | 262k | tool calling, reasoning controls | — |
| Google: Gemini 3.1 Pro Preview Listed Feb 19, 2026 Assessment is missing in the draft for Gemini 3.1 Pro Preview, so no score is provided here. | Limited evidence | $2.000 | $12.00 | 1M | image input, audio input, tool calling, reasoning controls, structured output | — | |
| Google: Gemini 3.1 Pro Preview (batch) Listed Feb 19, 2026 Assessment is missing in the draft for Gemini 3.1 Pro Preview, so no score is provided here. | Limited evidence | $1.000 | $6.000 | 1M | image input, audio input, tool calling, reasoning controls, structured output | — | |
| OpenAI: GPT-5 Image Mini Listed Oct 16, 2025 The draft assessment for GPT-5 Image Mini is truncated, so its scores and supporting detail are incomplete here. | OpenAI | Limited evidence | $2.500 | $2.000 | 400k | image input, reasoning controls, structured output | — |
| Qwen: Qwen3 Next 80B A3B Instruct Listed Sep 11, 2025 No assessment was supplied for this model, so capability cannot be scored from the provided draft. | Qwen | Limited evidence | $0.090 | $1.100 | 262k | tool calling, structured output | — |
| OpenAI: GPT-5 Nano Listed Aug 7, 2025 No assessment for this model was included in the supplied draft, so scores and conclusions are unavailable. | OpenAI | Limited evidence | $0.050 | $0.400 | 400k | image input, tool calling, reasoning controls, structured output | — |
| OpenAI: gpt-oss-20b Listed Aug 5, 2025 No assessment for gpt-oss-20b was provided in the draft, so there is not enough supplied evidence to score it. | OpenAI | Limited evidence | $0.018 | $0.090 | 131k | tool calling, reasoning controls, structured output | — |
| Qwen: Qwen3 235B A22B Thinking 2507 Listed Jul 25, 2025 No assessment for Qwen3-235B-A22B-Thinking-2507 is present in the supplied draft, so no scores can be reliably extracted. | Qwen | Limited evidence | $0.450 | $3.500 | 128k | tool calling, reasoning controls | — |
| Venice: Uncensored Listed Jul 9, 2025 No primary-source model card or technical report was available for this exact model ID, so there is not enough verified evidence to score it reliably. | Venice | Limited evidence | $0.200 | $0.900 | 128k | — | — |
| Tencent: Hunyuan A13B Instruct Listed Jul 8, 2025 The draft assessment is truncated, so there is not enough complete evidence here to score this model reliably. | Tencent | Limited evidence | $0.140 | $0.570 | 131k | reasoning controls, structured output | — |
| MiniMax: MiniMax M1 Listed Jun 17, 2025 No assessment for this model was included in the supplied draft, so scores and conclusions are unavailable. | MiniMax | Limited evidence | $0.550 | $2.200 | 1M | tool calling, reasoning controls | — |
| Qwen: Qwen3 32B Listed Apr 28, 2025 The supplied draft for Qwen3-32B is truncated, so the assessment cannot be completed reliably from the provided text. | Qwen | Limited evidence | $0.080 | $0.280 | 131k | tool calling, reasoning controls, structured output | — |
| OpenAI: GPT-4.1 Mini Listed Apr 14, 2025 No assessment for this model was included in the supplied draft, so scores and conclusions are unavailable. | OpenAI | Limited evidence AI review detailsLow confidence · Oct 17, 2026 Reviewed by gpt-5.4-mini. AI assessment of published evidence; not a hands-on benchmark. | $0.400 | $1.600 | 1M | image input, tool calling, structured output | — |
| OpenAI: GPT-4.1 Mini (batch) Listed Apr 14, 2025 No assessment for this batch variant was included in the supplied draft, so scores and conclusions are unavailable. | OpenAI | Limited evidence | $0.200 | $0.800 | 1M | image input, tool calling, structured output | — |
| OpenAI: o1-pro Listed Mar 19, 2025 The draft is truncated here, so capability claims for o1-pro are incomplete and should be treated cautiously. | OpenAI | Limited evidence | $150.00 | $600.00 | 200k | image input, reasoning controls, structured output | — |
| Google: Gemma 3 4B Listed Mar 13, 2025 No assessment for Gemma 3 4B IT was provided in the draft, so there is not enough supplied evidence to score it. | Limited evidence | $0.050 | $0.100 | 131k | image input, structured output | — |
Recent snapshot history
October 18, 2026
444 tracked models
First listed: DeepSeek: DeepSeek V4.1 Flash 1017, Elorian: Glimpse, XingChen: Xing4.0 29B A4B (free)
October 17, 2026
444 tracked models
First listed: DeepSeek: DeepSeek V4.1 Flash 1017, Elorian: Glimpse, XingChen: Xing4.0 29B A4B (free)
October 16, 2026
440 tracked models
First listed: XingChen: Xing4.0 29B A4B (free), Reflection: Beam 1.0 501B A23B, Moondream: Moondream 3 Preview
October 15, 2026
438 tracked models
First listed: Moondream: Moondream 3 Preview, Moondream: Moondream 3.1, Meta: Muse Spark Open
October 14, 2026
435 tracked models
First listed: Pareto 26.9, LiquidAI: LFM2.5-24B-A2B, Google: Gemini 4 Argon
October 13, 2026
435 tracked models
First listed: LiquidAI: LFM2.5-24B-A2B, Google: Gemini 4 Argon, Google: Gemini 4 Argon (batch)
October 12, 2026
433 tracked models
First listed: Meta: Muse Wonder, StepFun: Step 5 Preview, Anthropic: Claude Haiku 5.5