Skip to content
Model Index
  • Overview
  • Models
  • Compare
  • Benchmarks
  • Finder
LabsMethodology
  1. Benchmarks
  2. Agents & knowledge work

Arena · Agent

Human preferenceArena · updated Oct 2, 2026

What it measures

How well models orchestrate tools in real agent sessions, scored from behavioural signals such as task completion, user praise or complaints, steering burden, recovery steps and tool hallucinations.

How to read it

A relative score centred on the field: zero is average, positive is better. Based on live usage, so it reflects practice rather than a fixed test set.

Open at Arena
Best
+14.3
Claude Fable 5.1
Median
+2.4
GLM-5.3

Results

29 of 41 tracked models have a result. This scale has no meaningful zero, so bars run from the lowest to the highest tracked score.

Filter in explorer
  1. 1
    Claude Fable 5.1

    Anthropic · Max effort

    +14.3 ±1.9
  2. 2
    Claude Opus 5.5

    Anthropic · High effort

    +13.8 ±2.2
  3. 3
    Claude Sonnet 5.5

    Anthropic · Max effort

    +12.5 ±3.1
  4. 4
    GPT-6 Astra

    OpenAI · Max effort

    +12.3 ±2.2
  5. 5
    GPT-6.1 Sol

    OpenAI · Max effort

    +11.2 ±2.8
  6. 6
    Gemini 4 ArgonLimited

    Google DeepMind · High effort

    +7.6 ±1.9
  7. 7
    Kimi K3Open

    Moonshot AI · Max effort

    +4.2 ±0.5
  8. 8
    DeepSeek V4.1 FlashOpen

    DeepSeek · Max effort

    +4.0 ±0.5
  9. 8
    Grok 4.7

    SpaceXAI · XHigh effort

    +4.0 ±1.6
  10. 8
    Hy4 previewOpen

    Tencent

    +4.0 ±0.8
  11. 8
    Muse Spark 1.3

    Meta · Max effort

    +4.0 ±0.6
  12. 12
    MiMo-V2.6-ProOpen

    Xiaomi

    +3.3 ±1.7
  13. 13
    Gemini 3.8 Flash

    Google DeepMind · High effort

    +3.0 ±0.8
  14. 14
    Qwen3.8-Max

    Alibaba

    +2.5 ±0.6
  15. 15
    GLM-5.3Open

    Z.ai · Max effort

    +2.4 ±0.7
  16. 16
    GPT-6 Luna

    OpenAI · Max effort

    +1.4 ±1.1
  17. 17
    DeepSeek V4 ProOpen

    DeepSeek · High effort

    +1.2 ±1.7
  18. 18
    Step 5 Preview

    StepFun

    +0.5 ±1.5
  19. 19
    GPT-5.6 Terra

    OpenAI · XHigh effort

    +0.4 ±1.2
  20. 20
    GLM-5.3-FlashOpen

    Z.ai

    -0.4 ±0.5
  21. 21
    MiMo-V2.6-FlashOpen

    Xiaomi

    -0.6 ±1.3
  22. 22
    Qwen3.8-FlashOpen

    Alibaba

    -0.7 ±0.6
  23. 23
    Qwen3.8-27BOpen

    Alibaba

    -1.7 ±0.6
  24. 24
    Hy3Open

    Tencent

    -5.4 ±1.2
  25. 25
    MiniMax-M3Open

    MiniMax

    -6.9 ±0.9
  26. 26
    Gemini 3.1 Pro

    Google DeepMind

    -7.7 ±1
  27. 27
    TM
    Inkling-SmallOpen

    Thinking Machines Lab

    -10.3 ±1.4
  28. 28
    TM
    InklingOpen

    Thinking Machines Lab

    -10.9 ±1
  29. 29
    Mistral Medium 3.5Open

    Mistral AI

    -12.4 ±1.5
12 tracked models without a result

Claude Haiku 5.5, Command A+, Gemini 3.5 Flash-Lite, Gemma 4 31B, gpt-oss-120b, Kimi K2.7 Code, MAI-Thinking-1, Mistral Large 4, Mistral Small 4, Muse Glimmer 30B, Nemotron 3 Ultra, Nova 2 Lite.

Missing means the evaluator has not published a run, not that the model would score poorly.

Model Index

An independent snapshot of frontier SI models and how they perform on third-party benchmarks.

Data as of Oct 8, 2026

Explore

  • All models
  • Compare
  • Benchmarks
  • Model finder
  • Labs
  • Methodology & sources

Sources

Benchmark results come from Artificial Analysis, Arena, Vals AI and ARC Prize. Specifications and prices come from each lab's own documentation.

Not affiliated with any lab or evaluator. Names and logos are trademarks of their respective owners and are used only to identify them. Scores change often — check the linked source before making a decision that depends on one.

  • Overview
  • Models
  • Compare
  • Benchmarks
  • Finder
  • Overview
  • Models
  • Compare
  • Benchmarks
  • Finder
  • Overview
  • Models
  • Compare
  • Benchmarks
  • Finder