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 solve problems that stump experts across fields?Humanity's Last Exam
Claude Opus 5.5🥇
Claude Opus 5.5🥇Humanity's Last Exam
Claude Opus 5.5🥇Can it solve research-level physics problems?CritPt
Gemini 4 Carbon🥇
Gemini 4 Carbon🥇CritPt
Gemini 4 Carbon🥇Which AI gives the math answers people prefer most?Arena Math
Gemini 4 Argon🥇
Gemini 4 Argon🥇Arena Math
Gemini 4 Argon🥇
Can it write airtight math proofs?
ProofBench
- 🥇AnthropicClaude Fable 5.2Reasoning effort: Max100.0%
- 🥇AnthropicClaude Fable 5.1Reasoning effort: Max100.0%
- 🥇AnthropicClaude Opus 5.5Reasoning effort: Max100.0%
- 🥇AnthropicClaude Sonnet 5.5Reasoning effort: Max100.0%
- #5GoogleGemini 4 Argon99.0%
- #5OpenAIGPT-6.1 Sol99.0%
- #5AnthropicClaude Opus 5Reasoning effort: Max99.0%
- #5OpenAIGPT-6 Astra99.0%
- #9AnthropicClaude Fable 5Reasoning effort: Max95.0%
- #10Moonshot AIKimi K3.1 Pro91.0%
- #11AnthropicClaude Haiku 5.5Reasoning effort: Max89.0%
- #12Moonshot AIKimi K387.0%
- #13OpenAIGPT-6 Sol83.0%
- #13OpenAIGPT-5.6 SolReasoning effort: Max83.0%
- #15AnthropicClaude Sonnet 5Reasoning effort: Max77.0%
- #16TencentHy475.0%
- #17OpenAIGPT-5.6 TerraReasoning effort: Extra High74.0%
- #18XiaomiMiMo-V2.6-Pro70.0%
- #19AnthropicClaude Opus 4.8Reasoning effort: Max69.0%
- #20OpenAIGPT-6 Luna64.0%
- #21XiaomiMiMo-V2.6-Flash63.0%
- #22OpenAIGPT-5.6 LunaReasoning effort: Max60.0%
- #23MetaMuse Spark 1.3Reasoning effort: Max58.0%
- #23GoogleGemini 3.7 Flash58.0%
- #25DeepSeekDeepSeek V4 Flash56.0%
- #25OpenAIGPT-5.4Reasoning effort: Extra High56.0%
- #27DeepSeekDeepSeek V4.1 Flash54.0%
- #27AnthropicClaude Opus 4.7Reasoning effort: Max54.0%
- #29SpaceXSIGrok 4.651.0%
- #30AnthropicClaude Opus 4.6Reasoning effort: Max50.0%
- #30OpenAIGPT-5.5Reasoning effort: Extra High50.0%
- #30DeepSeekDeepSeek V4 Pro50.0%
- #33Z.aiGLM 5.3Reasoning effort: Max49.0%
- #34GoogleGemini 3.8 Flash48.0%
- #35AnthropicClaude Sonnet 4.6Reasoning effort: Max45.0%
- #36MetaMuse Spark 1.243.0%
- #37StepfunStep-542.0%
- #38MetaMuse Spark 1.139.0%
- #39AnthropicClaude Opus 4.536.0%
- #39GoogleGemini 3.6 Flash36.0%
- #41Z.aiGLM 5.2Reasoning effort: Max35.0%
- #42GoogleGemini 3.5 FlashReasoning effort: High31.0%
- #42SpaceXSIGrok 4.5Reasoning effort: High31.0%
- #44GoogleGemini 3.1 Pro26.0%
- #44AlibabaQwen3.7 Max26.0%
- #44SpaceXSIGrok 4.726.0%
- #47Z.aiGLM-5.122.2%
- #48XiaomiMiMo V2.5 Pro22.0%
- #49Z.aiGLM 5.3 FlashReasoning effort: Max21.0%
- #49OpenAIGPT-5.4 MiniReasoning effort: Extra High21.0%
- #51AnthropicClaude Sonnet 4.519.0%
- #52OpenAIGPT-5Reasoning effort: High18.0%
- #52MiniMaxMiniMax M318.0%
- #54MetaMuse Spark17.0%
- #55XiaomiMiMo V2.516.0%
- #55Moonshot AIKimi K2.616.0%
- #57GoogleGemini 3 Flash15.0%
- #58SpaceXSIGrok 4.20 (Reasoning)14.0%
- #59GoogleGemini 3.5 Flash-Lite13.0%
- #60OpenAIGPT-5 NanoReasoning effort: High12.0%
- #61SpaceXSIGrok 4.3Reasoning effort: High11.0%
- #62Mistral AIMistral Large 410.0%
- #63OpenAIGPT-5 MiniReasoning effort: High9.0%
- #63Mistral AIMistral Medium 3.59.0%
- #65DeepSeekDeepSeek V3.28.0%
- #66OpenAIGPT-5.4 NanoReasoning effort: High5.0%
- #67MiniMaxMiniMax M2.54.0%
- #67SpaceXSIGrok 4.1 Fast (Reasoning)4.0%
- #69MiniMaxMiniMax M2.73.0%
- #70NVIDIANemotron 3 Ultra2.0%
- #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.