ProofBench v1.1
Formal and semi-formal mathematical proof construction, graded on the proof rather than the final answer.
MathematicsVals AI46 models run by Vals AIUpdated 29 September 2026
How it is run
Several frontier models reach a perfect 100%, matching a specialised prover. Equal scores share a rank in the leaderboard.
What it cannot tell you
Saturated at the top for the frontier and steeply falling below it — useful for separating frontier from mid-tier, not for ordering the frontier.
- Leader
- Claude Sonnet 5.5100.00% · Vals AI's own list
- Top-ten spread
- 13.00%Leader minus tenth place, or last place in smaller fields.
- Median
- 57.00%of the 46 rows published
- No published result
- 12of 46 models on this site
Independent leaderboard · Vals AI
Top 46 of 46 runs by Vals AI, one harness.
| # | Model | Score |
|---|---|---|
| 1 | Claude Sonnet 5.5 | 100.00% |
| 1 | Claude Opus 5.5 | 100.00% |
| 1 | Claude Fable 5.1 | 100.00% |
| 1 | Alephprover | 100.00% |
| 5 | GPT-6.1 Sol | 99.00% |
| 5 | GPT-6 Astra | 99.00% |
| 5 | Gemini 4 Argon | 99.00% |
| 5 | Claude Opus 5 | 99.00% |
| 9 | Claude Fable 5 | 95.00% |
| 10 | Kimi K3 | 87.00% |
| 11 | Aristotle | 86.00% |
| 12 | GPT-6 Sol | 83.00% |
| 12 | GPT-5.6 Sol | 83.00% |
| 14 | Claude Sonnet 5 | 77.00% |
| 15 | Hunyuan Hy4 Preview | 75.00% |
| 16 | GPT-5.6 Terra | 74.00% |
| 17 | MiMo V2.6 Pro | 70.00% |
| 18 | GPT-6 Luna | 64.00% |
| 19 | MiMo-V2.6-Flash | 63.00% |
| 20 | GPT-5.6 Luna | 60.00% |
| 21 | Gemini 3.7 Flash | 58.00% |
| 21 | Muse Spark 1.3 Max | 58.00% |
| 21 | Qwen3.8 Max | 58.00% |
| 24 | Deepseek V4 Flash 0731 | 56.00% |
| 25 | Muse Spark 1 3 | 55.00% |
| 26 | DeepSeek V4.1 Flash | 54.00% |
| 27 | Grok 4.6 | 51.00% |
| 28 | DeepSeek V4 Pro | 50.00% |
| 29 | GLM 5.3 | 49.00% |
| 30 | Gemini 3.8 Flash | 48.00% |
| 31 | Muse Spark 1.2 | 43.00% |
| 32 | Gemini 3.5 Flash | 31.00% |
| 32 | Grok 4.5 | 31.00% |
| 34 | Grok 4.7 | 26.00% |
| 34 | Gemini 3.1 Pro | 26.00% |
| 36 | Mimo V2.5 Pro | 22.00% |
| 37 | GLM 5.3 Flash | 21.00% |
| 38 | MiniMax M3 | 18.00% |
| 39 | Deepseek V4 Pro | 16.00% |
| 39 | Mimo V2.5 | 16.00% |
| 39 | Qwen3.8 27b | 16.00% |
| 42 | Mistral Medium 3.5 | 9.00% |
| 43 | Inkling Small | 6.00% |
| 44 | Mercury 2.5 | 3.00% |
| 45 | Inkling | 0.00% |
| 45 | Nemotron Lightning 3p5 30b A3b | 0.00% |
No published result tracked
12 of the 46 models on this site have no published number on this benchmark, so their cells in the home table stay blank rather than reading as zero.
Read this with
Benchmarks in the same category, which usually disagree with this one on purpose.
Sources
Every page used for this benchmark, read on 17 October 2026.
- Benchmark site or the page that publishes itwww.vals.ai/benchmarks/proof_bench
- Independent resultswww.vals.ai/benchmarks/proof_bench