AI rankings for documents and office work

APEX-Agents

Often citedHigher is better
What does it measure?
Tests whether an AI can carry out professional tasks from investment banking, management consulting and corporate law inside simulated work environments. Made by Mercor.
Who evaluates it, and how?
480 tasks, 160 per profession, done in 33 simulated work environments with documents, spreadsheets, email and calendars (no web search). A task passes only if every expert-written rubric item is met; the score is the first-attempt pass rate.
Reading the score
Scores run from 0 to 100%. Higher is better.
Top score84.2%Gemini 4 Carbon
Models tested54
Model release date (newest on the right)
Best score so farHigher on the chart is better
  1. 🥇GoogleGemini 4 Carbon84.2%
  2. 🥈GoogleGemini 4 Argon82.4%
  3. 🥉AnthropicClaude Fable 5.280.1%
  4. #4AnthropicClaude Sonnet 5.5Reasoning effort: Max75.5%
  5. #5AnthropicClaude Opus 5.5Reasoning effort: Max73.5%
  6. #6DeepSeekDeepSeek V4.1 Flash 1017Reasoning effort: Max69.8%
  7. #7AnthropicClaude Fable 5.168.6%
  8. #8OpenAIGPT-6.1 TerraReasoning effort: Max68.3%
  9. #9GoogleGemini 3.7 Flash67.8%
  10. #10AnthropicClaude Opus 5Reasoning effort: Max65.8%
  11. #11Moonshot AIKimi K3.1 Pro65.5%
  12. #12SpaceXSIGrok 4.665.3%
  13. #13OpenAIGPT-6 Astra64.7%
  14. #14GoogleGemini 3.8 Flash64.3%
  15. #14MetaMuse Ember64.3%
  16. #16AnthropicClaude Fable 563.6%
  17. #17OpenAIGPT-6.1 SolReasoning effort: Max60.0%
  18. #18MetaMuse Spark 1.358.6%
  19. #19OpenAIGPT-5.6 TerraReasoning effort: Max58.2%
  20. #20Z.aiGLM 5.356.6%
  21. #21SpaceXSIGrok 4.556.2%
  22. #22AnthropicClaude Haiku 5.5Reasoning effort: Max56.1%
  23. #23OpenAIGPT-5.555.1%
  24. #24SpaceXSIGrok 4.754.6%
  25. #25AnthropicClaude Sonnet 554.5%
  26. #26XiaomiMiMo-V2.6-Pro54.4%
  27. #27OpenAIGPT-6 SolReasoning effort: Max54.3%
  28. #28Z.aiGLM 5.3 Flash52.8%
  29. #29OpenAIGPT-5.452.4%
  30. #30Mistral AIMistral Large 4Reasoning effort: High51.6%
  31. #31OpenAIGPT-5.6 SolReasoning effort: Max51.4%
  32. #32AnthropicClaude Opus 4.7Reasoning effort: Max49.2%
  33. #33AnthropicClaude Opus 4.8Reasoning effort: Max48.9%
  34. #34Moonshot AIKimi K347.8%
  35. #35DeepSeekDeepSeek V4 Pro47.3%
  36. #36GoogleGemini 3.6 Flash46.9%
  37. #37AnthropicClaude Opus 4.6Reasoning effort: Max46.3%
  38. #38MetaMuse Spark 1.245.4%
  39. #39Z.aiGLM 5.245.2%
  40. #40OpenAIGPT-6 LunaReasoning effort: Max44.3%
  41. #41AnthropicClaude Sonnet 4.6Reasoning effort: High43.0%
  42. #41OpenAIGPT-5.6 Luna43.0%
  43. #43Z.aiGLM-5.140.9%
  44. #44DeepSeekDeepSeek V4.1 Flash39.5%
  45. #45MiniMaxMiniMax M337.7%
  46. #46Moonshot AIKimi K2.7 Code37.6%
  47. #47GoogleGemini 3.1 Pro35.1%
  48. #48MetaMuse Spark 1.131.8%
  49. #49GoogleGemini 3.5 Flash-Lite29.3%
  50. #50GoogleGemini 3.5 Flash27.5%
  51. #51AlibabaQwen3.5 397B A17B24.9%
  52. #52NVIDIANemotron 3 Ultra22.7%
  53. #53DeepSeekDeepSeek V3.221.3%
  54. #54OpenAIGPT OSS 120B4.4%
Half of models ≤ 53.6%Last updated 2026-11-11

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.

Where the scores come from

  • Run by the test’s team: scores the team that operates the test ran and published itself.