ProofBench (Vals)

What does it measure?
Tests whether a model can write proofs for graduate and advanced-undergraduate level theorems. Made by Vals AI.
Who evaluates it, and how?
For 100 theorems in analysis, algebra, probability, number theory, and more, the model writes a proof in Lean 4, a proof-checking language, and it must pass the machine checker. There is no partial credit.
Reading the score
Scores run from 0 to 100%. Higher is better. It is not the same as Google DeepMind's similarly named test.
Top score100.0%
Models tested71
Scoring basisHigher is better
Model release date

Can it write airtight math proofs?

ProofBench

  1. 🥇AnthropicClaude Fable 5.2Reasoning effort: Max100.0%
  2. 🥇AnthropicClaude Fable 5.1Reasoning effort: Max100.0%
  3. 🥇AnthropicClaude Opus 5.5Reasoning effort: Max100.0%
  4. 🥇AnthropicClaude Sonnet 5.5Reasoning effort: Max100.0%
  5. #5GoogleGemini 4 Argon99.0%
  6. #5OpenAIGPT-6.1 Sol99.0%
  7. #5AnthropicClaude Opus 5Reasoning effort: Max99.0%
  8. #5OpenAIGPT-6 Astra99.0%
  9. #9AnthropicClaude Fable 5Reasoning effort: Max95.0%
  10. #10Moonshot AIKimi K3.1 Pro91.0%
  11. #11AnthropicClaude Haiku 5.5Reasoning effort: Max89.0%
  12. #12Moonshot AIKimi K387.0%
  13. #13OpenAIGPT-6 Sol83.0%
  14. #13OpenAIGPT-5.6 SolReasoning effort: Max83.0%
  15. #15AnthropicClaude Sonnet 5Reasoning effort: Max77.0%
  16. #16TencentHy475.0%
  17. #17OpenAIGPT-5.6 TerraReasoning effort: Extra High74.0%
  18. #18XiaomiMiMo-V2.6-Pro70.0%
  19. #19AnthropicClaude Opus 4.8Reasoning effort: Max69.0%
  20. #20OpenAIGPT-6 Luna64.0%
  21. #21XiaomiMiMo-V2.6-Flash63.0%
  22. #22OpenAIGPT-5.6 LunaReasoning effort: Max60.0%
  23. #23MetaMuse Spark 1.3Reasoning effort: Max58.0%
  24. #23GoogleGemini 3.7 Flash58.0%
  25. #25DeepSeekDeepSeek V4 Flash56.0%
  26. #25OpenAIGPT-5.4Reasoning effort: Extra High56.0%
  27. #27DeepSeekDeepSeek V4.1 Flash54.0%
  28. #27AnthropicClaude Opus 4.7Reasoning effort: Max54.0%
  29. #29SpaceXSIGrok 4.651.0%
  30. #30AnthropicClaude Opus 4.6Reasoning effort: Max50.0%
  31. #30OpenAIGPT-5.5Reasoning effort: Extra High50.0%
  32. #30DeepSeekDeepSeek V4 Pro50.0%
  33. #33Z.aiGLM 5.3Reasoning effort: Max49.0%
  34. #34GoogleGemini 3.8 Flash48.0%
  35. #35AnthropicClaude Sonnet 4.6Reasoning effort: Max45.0%
  36. #36MetaMuse Spark 1.243.0%
  37. #37StepfunStep-542.0%
  38. #38MetaMuse Spark 1.139.0%
  39. #39AnthropicClaude Opus 4.536.0%
  40. #39GoogleGemini 3.6 Flash36.0%
  41. #41Z.aiGLM 5.2Reasoning effort: Max35.0%
  42. #42GoogleGemini 3.5 FlashReasoning effort: High31.0%
  43. #42SpaceXSIGrok 4.5Reasoning effort: High31.0%
  44. #44GoogleGemini 3.1 Pro26.0%
  45. #44AlibabaQwen3.7 Max26.0%
  46. #44SpaceXSIGrok 4.726.0%
  47. #47Z.aiGLM-5.122.2%
  48. #48XiaomiMiMo V2.5 Pro22.0%
  49. #49Z.aiGLM 5.3 FlashReasoning effort: Max21.0%
  50. #49OpenAIGPT-5.4 MiniReasoning effort: Extra High21.0%
  51. #51AnthropicClaude Sonnet 4.519.0%
  52. #52OpenAIGPT-5Reasoning effort: High18.0%
  53. #52MiniMaxMiniMax M318.0%
  54. #54MetaMuse Spark17.0%
  55. #55XiaomiMiMo V2.516.0%
  56. #55Moonshot AIKimi K2.616.0%
  57. #57GoogleGemini 3 Flash15.0%
  58. #58SpaceXSIGrok 4.20 (Reasoning)14.0%
  59. #59GoogleGemini 3.5 Flash-Lite13.0%
  60. #60OpenAIGPT-5 NanoReasoning effort: High12.0%
  61. #61SpaceXSIGrok 4.3Reasoning effort: High11.0%
  62. #62Mistral AIMistral Large 410.0%
  63. #63OpenAIGPT-5 MiniReasoning effort: High9.0%
  64. #63Mistral AIMistral Medium 3.59.0%
  65. #65DeepSeekDeepSeek V3.28.0%
  66. #66OpenAIGPT-5.4 NanoReasoning effort: High5.0%
  67. #67MiniMaxMiniMax M2.54.0%
  68. #67SpaceXSIGrok 4.1 Fast (Reasoning)4.0%
  69. #69MiniMaxMiniMax M2.73.0%
  70. #70NVIDIANemotron 3 Ultra2.0%
  71. #71PoolsideLaguna M.10.0%
Top 4 within the margin of error · effectively tied · Half of models ≤ 43.0%

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.