Browse 34,691 LLMs

A long-horizon repository benchmark built to test realistic software engineering work. Its scores need a task-quality and setup check before they support a coding-agent decision.

1. 1
   
   Claude Opus 5.589.9
2. 2
   
   Claude Sonnet 5.581.3
3. 3
   
   Claude Fable 5.181.2

85 models

A multilingual software-engineering benchmark for real-world code issue resolution across multiple programming languages.

1. 1
   
   Claude Opus 5.593.9
2. 2
   
   Claude Sonnet 5.590.3
3. 3
   
   Claude Opus 589.5

60 models

An independently evaluated Tau3 banking benchmark from Artificial Analysis.

1. 1
   
   Grok 4.650.7
2. 2
   
   Muse Spark 1.350.5
3. 3
   
   GLM-5.350.3

22 models

How important are skills for agents?

1. 1
   
   DeepSeek V4.1 Flash 101769.8
2. 2
   
   Grok 4.566.0
3. 3
   
   Gemini 3.7 Flash65.9

35 models

A broad research-agent benchmark for open-ended information gathering, synthesis, and answer construction across wide search spaces.

1. 1
   Me
   Mercury 292.3
2. 2
   Hy
   Hy4 preview83.9
3. 3
   Li
   Ling 3.1 Flash83.3

17 models

BrowseComp tests whether a browsing agent can find difficult, fact-seeking answers across the web. Its 1,266 problems use short answers that are easier to grade than an open-ended research report.

1. 1
   At
   Atria Dawn Preview92.5
2. 2
   
   GPT-5.6 Sol92.2
3. 3
   
   Kimi K3.1 Pro92.1

62 models

MRCR v2 slice focused on very long contexts at 128K-256K lengths.

1. 1
   
   GPT-5.587.5
2. 2
   
   Claude Opus 4.7 (Adaptive)59.2

2 models

Human-validated software engineering issues from real-world Python repositories.

1. 1
   
   Claude Opus 4.7 (max)83.5
2. 2
   
   GPT-5.5 (xhigh)80.6
3. 3
   
   Gemini 3.5 Flash (high)79.3

33 models

Factoid questions spanning politics, science, technology, art, sports, geography, and music.

1. 1
   
   Gemini 4 Argon (high)76.0
2. 2
   
   GPT-6 Astra (max)75.6
3. 3
   
   GPT-6.1 Sol (max)73.9

93 models

Abstract reasoning and pattern generalization on grid-based tasks.

1. 1
   
   GPT-6 Astra (max)95.0
2. 2
   
   GPT-6.1 Sol (max)94.2
3. 3
   
   GPT-6 Astra (xhigh)93.3

266 models

Exceptionally difficult research-level mathematics problems.

1. 1
   
   GPT-6.1 Sol (max)100.0
2. 2
   
   GPT-6 Astra (high)97.6
3. 3
   
   GPT-6 Astra (xhigh)97.6

78 models

Real-world multi-step tool-use evaluation through the Model Context Protocol.

1. 1
   
   Muse Spark 1.188.1
2. 2
   Fa
   Fable 5.187.2
3. 3
   
   claude-opus-5 (xhigh)85.8

34 models

Real-world cybersecurity evaluation of AI agents reproducing vulnerabilities with working proof-of-concept tests.

1. 1
   
   DeepSeek-V4.1-Flash99.4
2. 2
   
   Multi-model (DeepSeek-V4.1-Flash, Abliterated GLM-5.2)99.2
3. 3
   
   Multi-model (Creation Model, DeepSeek-V4-Pro, Qwen 3.8 Max)98.5

86 models

The official leaderboard for Terminal-Bench 3.0.

1. 1
   Op
   Opus 542.7
2. 2
   
   GPT-5.6 Sol34.6
3. 3
   Fa
   Fable 534.0

12 models

Expert-level questions across mathematics, science, and humanities, as published by the CAIS AI Dashboard.

1. 1
   
   Gemini 4 Argon55.6
2. 2
   Op
   Opus 5.555.4
3. 3
   Fa
   Fable 5.154.6

66 models

Average of the text capability benchmarks published by the CAIS AI Dashboard.

1. 1
   
   Gemini 4 Argon66.5
2. 2
   
   GPT-6 Astra63.9
3. 3
   
   GPT-6.1 Sol61.4

66 models

Average of the vision capability benchmarks published by the CAIS AI Dashboard.

1. 1
   
   Gemini 4 Argon83.3
2. 2
   
   GPT-6 Astra83.0
3. 3
   
   GPT-6.1 Sol77.9

53 models

Average of the risk and safety benchmarks published by the CAIS AI Dashboard; lower is better.

1. 1
   Op
   Opus 5.526.2
2. 2
   Fa
   Fable 5.130.3
3. 3
   
   GPT-6 Astra35.5

11 models

Remote Labor Index automation rates published by the CAIS AI Dashboard.

1. 1
   Op
   Opus 5.521.3
2. 2
   
   GPT-6 Astra20.8
3. 3
   Fa
   Fable 5.117.9

18 models

Composite Artificial Analysis index across mathematics, science, coding, and reasoning evaluations.

1. 1
   
   Gemini 4 Carbon (high)59.4
2. 2
   
   Claude Opus 5.5 (max with fallback)58.4
3. 3
   
   Claude Opus 5.5 (xhigh with fallback)57.0

86 models

Artificial Analysis composite index for programming and software-engineering capability.

1. 1
   
   Claude Fable 5.1 (max with fallback)81.6
2. 2
   
   Claude Fable 5.1 (xhigh with fallback)80.7
3. 3
   
   Claude Fable 5.1 (high with fallback)79.1

256 models

Artificial Analysis factual-knowledge and hallucination-resistance evaluation.

1. 1
   
   Gemini 4 Carbon (high)53.5
2. 2
   
   Claude Opus 5.5 (max with fallback)48.1
3. 3
   
   Claude Fable 5.1 (max with fallback)46.3

59 models

Artificial Analysis composite index for tool use, planning, autonomy, and complex agentic workflows.

1. 1
   
   Claude Fable 5.1 (max with fallback)58.0
2. 2
   
   Claude Fable 5.1 (xhigh with fallback)57.1
3. 3
   
   Claude Opus 5 (max)56.2

151 models

Graduate-level, expert-written science questions evaluated by Artificial Analysis.

1. 1
   
   GPT-6 Astra (xhigh)96.3
2. 2
   
   GPT-6 Astra (max)96.1
3. 3
   
   Gemini 3.8 Flash (high)95.3

210 models

Broad expert-level reasoning and knowledge evaluation published by Artificial Analysis.

1. 1
   
   Claude Opus 5.5 (max with fallback)61.4
2. 2
   
   Claude Fable 5.1 (max with fallback)59.1
3. 3
   
   Claude Fable 5.1 (xhigh with fallback)58.7

265 models

Instruction-following benchmark results published by Artificial Analysis.

1. 1
   
   MiniMax-M382.9
2. 2
   Ne
   Nemotron 3 Ultra81.4
3. 3
   Ne
   Nemotron Cascade 2 30B A3B80.4

150 models

Scientific programming benchmark results published by Artificial Analysis.

1. 1
   
   Claude Opus 5.5 (max with fallback)66.9
2. 2
   
   Claude Opus 5.5 (xhigh with fallback)65.0
3. 3
   
   Claude Fable 5.1 (max with fallback)63.1

140 models

Agentic terminal and software-engineering results published by Artificial Analysis.

1. 1
   
   Claude Fable 5.1 (max with fallback)91.4
2. 2
   
   Claude Fable 5.1 (xhigh with fallback)91.0
3. 3
   
   Claude Fable 5.1 (high with fallback)89.9

118 models

Critical-points programming evaluation published by Artificial Analysis.

1. 1
   
   Muse Wonder Full (max)32.9
2. 2
   
   GPT-6.1 Sol (xhigh)31.7
3. 3
   
   GPT-6 Astra (max)31.7

264 models

Long-context reasoning results published by Artificial Analysis.

1. 1
   St
   Step 5 Preview88.3
2. 2
   
   MiMo-V2.6-Pro86.3
3. 3
   
   Kimi K3.1 Pro86.3

257 models

Tool-agent interaction benchmark results published by Artificial Analysis.

1. 1
   JT
   JT-35B-Flash99.1
2. 2
   
   Qwen3.5 397B A17B95.6
3. 3
   
   Gemini 3.1 Pro Preview95.6

147 models