ProofBench
Integrity rank #37 of 65 · 69 models scored · top score 100 · Claude Fable 5.1
Formal theorem-proving capability on graduate-level mathematics problems formalized in Lean 4.
Strongest on discrimination, weakest on contamination resistance. Caveat: the public test set does not rule out training-data contamination.
Reported scores — protocols may differ. How we compare
| # | Model | Score | Evidence | Measured |
|---|---|---|---|---|
| 1 | Claude Fable 5.1 | 100 | unverifiedT1 | 2026-09-01 |
| 2 | Claude Opus 5.5 | 100 | unverifiedT1 | 2026-09-22 |
| 3 | Claude Sonnet 5.5 | 100 | unverifiedT1 | 2026-09-28 |
| 4 | GPT-6.1 Sol | 99 | unverifiedT1 | 2026-09-29 |
| 5 | Gemini 4 Argon | 99 | unverifiedT1 | 2026-09-30 |
| 6 | Claude Opus 5 | 99 | unverifiedT1 | 2026-07-24 |
| 7 | GPT-6 Astra | 99 | unverifiedT1 | 2026-09-03 |
| 8 | Claude Fable 5 | 95 | unverifiedT1 | 2026-06-09 |
| 9 | Claude Haiku 5.5 | 89 | unverifiedT1 | 2026-10-07 |
| 10 | Kimi K3 | 87 | unverifiedT1 | 2026-07-16 |
| 11 | GPT-6 Sol | 83 | unverifiedT1 | 2026-09-22 |
| 12 | GPT-5.6 Sol | 83 | unverifiedT1 | 2026-07-09 |
| 13 | Claude Sonnet 5 | 77 | unverifiedT1 | 2026-06-30 |
| 14 | GPT-5.6 Terra | 74 | unverifiedT1 | 2026-07-09 |
| 15 | Claude Opus 4.8 | 69 | unverifiedT1 | 2026-05-28 |
| 16 | GPT-6 Luna | 64 | unverifiedT1 | 2026-09-22 |
| 17 | GPT-5.6 Luna | 60 | unverifiedT1 | 2026-07-09 |
| 18 | Gemini 3.7 Flash | 58 | unverifiedT1 | 2026-08-13 |
| 19 | Muse Spark 1.3 | 58 | unverifiedT1 | 2026-09-02 |
| 20 | Qwen 3.8 Max | 58 | unverifiedT1 | 2026-08-02 |
| 21 | DeepSeek V4 Flash 0731 | 56 | unverifiedT1 | 2026-07-31 |
| 22 | GPT-5.4 | 56 | unverifiedT1 | 2026-03-05 |
| 23 | Claude Opus 4.7 | 54 | unverifiedT1 | 2026-04-16 |
| 24 | DeepSeek V4.1 Flash | 54 | unverifiedT1 | 2026-09-09 |
| 25 | Grok 4.6 | 51 | unverifiedT1 | 2026-08-12 |
| 26 | Claude Opus 4.6 | 50 | unverifiedT1 | 2026-02-05 |
| 27 | DeepSeek V4 Pro 0813 | 50 | unverifiedT1 | 2026-08-13 |
| 28 | GPT-5.5 | 50 | unverifiedT1 | 2026-04-23 |
| 29 | GLM-5.3 | 49 | unverifiedT1 | 2026-08-14 |
| 30 | Gemini 3.8 Flash | 48 | unverifiedT1 | 2026-09-02 |
| 31 | Claude Sonnet 4.6 | 45 | unverifiedT1 | 2026-02-17 |
| 32 | Muse Spark 1.2 | 43 | unverifiedT1 | 2026-08-05 |
| 33 | Muse Spark 1.1 | 39 | unverifiedT1 | 2026-07-09 |
| 34 | Gemini 3.6 Flash | 36 | unverifiedT1 | 2026-07-21 |
| 35 | Claude Opus 4.5 | 36 | unverifiedT1 | 2025-11-24 |
| 36 | GLM-5.2 | 35 | unverifiedT1 | 2026-06-16 |
| 37 | Grok 4.7 | 34 | unverifiedT1 | 2026-09-21 |
| 38 | Gemini 3.5 Flash | 31 | unverifiedT1 | 2026-05-19 |
| 39 | Grok 4.5 | 31 | unverifiedT1 | 2026-07-08 |
| 40 | Gemini 3.1 Pro | 26 | unverifiedT1 | 2026-02-19 |
| 41 | Qwen3.7-Max | 26 | unverifiedT1 | 2026-05-19 |
| 42 | GLM-5.1 | 22.22 | unverifiedT1 | 2026-04-07 |
| 43 | GLM-5.3-Flash | 21 | unverifiedT1 | 2026-08-20 |
| 44 | GPT-5.4 Mini | 21 | unverifiedT1 | 2026-03-17 |
| 45 | Gemini 3 Pro | 20 | unverifiedT1 | 2025-11-18 |
| 46 | Claude Sonnet 4.5 | 19 | unverifiedT1 | 2025-09-29 |
| 47 | MiniMax-M3 | 18 | unverifiedT1 | 2026-06-01 |
| 48 | GPT-5 | 18 | unverifiedT1 | 2025-08-07 |
| 49 | Muse Spark | 17 | unverifiedT1 | 2026-04-08 |
| 50 | DeepSeek-V4-Pro | 16 | unverifiedT1 | 2026-04-24 |
| 51 | Kimi K2.6 | 16 | unverifiedT1 | 2026-04-20 |
| 52 | Qwen 3.8 27B | 16 | unverifiedT1 | 2026-08-14 |
| 53 | Gemini 3 Flash | 15 | unverifiedT1 | 2025-12-17 |
| 54 | GPT-5.2 | 15 | unverifiedT1 | 2025-12-11 |
| 55 | Grok 4.20 | 14 | unverifiedT1 | 2026-02-17 |
| 56 | Gemini 3.5 Flash-Lite | 13 | unverifiedT1 | 2026-07-21 |
| 57 | GPT-5 nano | 12 | unverifiedT1 | 2025-08-07 |
| 58 | Grok 4.3 Beta | 11 | unverifiedT1 | 2026-04-17 |
| 59 | Mistral Large 4 | 10 | unverifiedT1 | 2026-10-06 |
| 60 | Mistral Medium 3.5 | 9 | unverifiedT1 | 2026-04-28 |
| 61 | GPT-5 mini | 9 | unverifiedT1 | 2025-08-07 |
| 62 | DeepSeek-V3.2 | 8 | unverifiedT1 | 2025-12-01 |
| 63 | Inkling-Small | 6 | unverifiedT1 | 2026-07-15 |
| 64 | GLM-4.7 | 6 | unverifiedT1 | 2025-12-22 |
| 65 | GPT-5.4 Nano | 5 | unverifiedT1 | 2026-03-17 |
| 66 | MiniMax-M2.5 | 4 | unverifiedT1 | 2026-02-12 |
| 67 | MiniMax-M2.7 | 3 | unverifiedT1 | 2026-03-18 |
| 68 | Nemotron 3 Ultra | 2 | unverifiedT1 | 2026-06-04 |
| 69 | Inkling | 0 | unverifiedT1 | 2026-07-15 |
Continue from this leaderboard into the tightest head-to-head reads.
Weighted composite. Each dimension opens to the source-backed sub-signals it is computed from.
Claims drawn from cited facts, not live model generation.
This benchmark tests formal theorem proving on graduate-level mathematics problems.
1 cited fact
This benchmark has scores from 69 tracked models, with a recorded score range spanning 100.0.
2 cited facts
Its Benchmark Integrity Index score is 84, with grade B and rank 37 out of 65 tracked benchmarks.
4 cited facts
It is flagged as saturated. The top recorded score is 100.0, with 0.0 points of headroom. The computed top-score cluster contains 7 models.
4 cited facts
The harness classification is consistent harness. Test-set privacy is recorded as unknown. The recorded contamination-window ratio is 0.855.
3 cited facts
ProofBench: ProofBench as reported in Epoch AI's Capabilities Index CSV.
Claude Fable 5.1 leads ProofBench at 100. The full leaderboard above lists every recorded measurement, not just the headline number.
69 models have recorded scores on ProofBench, spanning a score spread of 100.
tensor.news grades ProofBench B for integrity (score 84/100), ranking #37 of 65 benchmarks we assess across discrimination, saturation, contamination resistance, harness comparability, and freshness.
Yes — top scores are clustering near the ceiling (saturation ratio 0.10), so ProofBench no longer separates leading models well. Treat small gaps at the top with caution.
Its test set is unknown, so contamination risk is on the table: a high score may partly reflect training-data overlap rather than capability.
Scores are reported under a consistent harness, so comparisons on ProofBench are reasonably apples-to-apples.
Every ProofBench measurement is source-backed and tagged reproduced, self-reported, or unverified; the underlying record cites www.vals.ai/benchmarks/proof_bench.