MCP-Atlas

Citation frequency 2/3
What does it measure?
A Scale AI test of whether the AI can complete tasks using real tools connected through MCP, a standard for linking AI to external tools.
Who evaluates it, and how?
1,000 tasks using 36 tool servers, 500 of which are public. Scale's own leaderboard is graded by an AI judge and limits each run to 100 tool calls.
Reading the score
Scores run from 0 to 100%. Higher is better. Values here were run by the model makers with different AI judges, so they can differ from Scale's leaderboard.

This ranking is built only from scores the model makers published themselves. Test settings such as tool use and reasoning effort differ by maker, so check each score's setting before comparing models directly.

Top score90.3%
Models tested37
Scoring basisHigher is better
Last updated2026.09.20
Model release date

Can it pick the right tools and use them properly?

MCP-Atlas

  1. 🥇MetaMuse Spark 1.2Reasoning effort: Extra High90.3%
  2. 🥈MetaMuse Spark 1.1Reasoning effort: Extra High88.1%
  3. 🥉AnthropicClaude Opus 585.8%
  4. #4StepfunStep-5Reasoning effort: High85.6%
  5. #5Moonshot AIKimi K3Reasoning effort: Max84.2%
  6. #6TencentHy4Reasoning effort: Max83.7%
  7. #7GoogleGemini 3.5 Flash83.6%
  8. #8AnthropicClaude Fable 583.3%
  9. #9AnthropicClaude Opus 4.8Reasoning effort: Max82.2%
  10. #9MetaMuse Spark82.2%
  11. #11AnthropicClaude Opus 4.7Reasoning effort: High79.7%
  12. #12GoogleGemini 3.1 Pro78.2%
  13. #13Z.aiGLM 5.276.8%
  14. #14AlibabaQwen3.7 Max76.4%
  15. #15Moonshot AIKimi K2.7 CodeReasoning effort: Thinking76.0%
  16. #16AnthropicClaude Opus 4.675.8%
  17. #17OpenAIGPT-5.5Reasoning effort: Extra High75.3%
  18. #18MiniMaxMiniMax M374.2%
  19. #18DeepSeekDeepSeek V4 ProReasoning effort: High74.2%
  20. #20AlibabaQwen3.6 Plus74.1%
  21. #21AlibabaQwen3.7 Plus73.2%
  22. #22Z.aiGLM-5.171.8%
  23. #23OpenAIGPT-5.4Reasoning effort: Extra High70.6%
  24. #24Moonshot AIKimi K2.6Reasoning effort: Thinking69.4%
  25. #25DeepSeekDeepSeek V4 FlashReasoning effort: Max69.0%
  26. #26Z.aiGLM-5Reasoning effort: Thinking67.8%
  27. #27AlibabaQwen3.6 Flash62.8%
  28. #27AlibabaQwen3.6 35B A3B62.8%
  29. #29AnthropicClaude Opus 4.5Reasoning effort: None62.3%
  30. #30GoogleGemini 3 Flash62.0%
  31. #31UpstageSolar Pro 461.4%
  32. #32AnthropicClaude Sonnet 4.6Reasoning effort: Max61.3%
  33. #33OpenAIGPT-5.4 MiniReasoning effort: Extra High57.7%
  34. #34OpenAIGPT-5.4 NanoReasoning effort: Extra High56.1%
  35. #35MiniMaxMiniMax M2.749.4%
  36. #36AnthropicClaude Sonnet 4.543.8%
  37. #37GoogleGemini 2.5 ProReasoning effort: Thinking8.8%
Half of models ≤ 74.2%Measured by: Maker-reported

The "harness" is the agent program the AI used to carry out the task. The same model can score very differently depending on its harness and reasoning effort.

  • Self-reported by each model maker — see the source link on each scoreFigures that model makers announced themselves, checked against the original post or system card and copied over. Conditions differ by maker, and each score links to its original source.