ProofBench: leaderboard
A mirror of the Vals AI ProofBench leaderboard as republished by benchmarklist.com, from a snapshot dated 28 July 2026. Upstream a model is given a theorem in natural language plus its Lean 4 formalisation and has to produce a proof the Lean checker accepts, using search and execution tools. The figure is the percentage of problems proved, higher is better. Source: www.vals.ai/benchmarks/proof_bench.
Metric: Score (imported). Source: benchmarklist.com. Source: epoch.ai. Saturation forecast: Around August 2027. 54 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5 | 78 | #12 |
| 2 | GPT-5.6 Sol (Max) | 77 | #19 (GPT-5.6 Sol) |
| 3 | Claude Fable 5 (Max) | 77 | #16 (Claude Fable 5) |
| 4 | GPT-5.6 Terra (xHigh) | 71 | #41 (GPT-5.6 Terra) |
| 5 | Kimi K3 | 70 | #30 |
| 6 | Claude Opus 4.8 (Max) | 69 | #37 (Claude Opus 4.8) |
| 7 | Claude Sonnet 5 (Max) | 66 | #57 (Claude Sonnet 5) |
| 8 | GPT-5.4 (2026-03-05) (xHigh) | 56 | #59 (GPT-5.4 (2026-03-05)) |
| 9 | GPT-5.6 Luna (Max) | 54 | #68 (GPT-5.6 Luna) |
| 10 | Claude Opus 4.7 (Max) | 54 | #49 (Claude Opus 4.7) |
| 11 | GPT-5.5 (xHigh) | 50 | #31 (GPT-5.5) |
| 12 | Claude Opus 4.6 (Max) | 50 | #63 (Claude Opus 4.6) |
| 13 | Claude Opus 4.6 (Adaptive Reasoning, Max Effort) | 50 | #63 (Claude Opus 4.6) |
| 14 | Claude Sonnet 4.6 (Max) | 45 | #90 (Claude Sonnet 4.6) |
| 15 | Muse Spark 1.1 (xHigh) | 39 | #52 (Muse Spark 1.1) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=proofbench · How It Works · Data refreshed daily, snapshot 2026-10-20.