Vals AI ProofBench: leaderboard
Each of 100 problems pairs a natural-language theorem with a vetted Lean 4 statement, across seven areas including measure theory, algebraic geometry and model theory. The model works for up to 40 turns with Mathlib search and Lean execution, then submits once, and the proof counts only if the Lean checker accepts it, with no partial credit. The number is the percentage of problems proved, higher is better.
Metric: Accuracy (%). Source: www.vals.ai. Source: epoch.ai. Saturation forecast: Estimated already saturated. 52 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Fable 5.1 (Max) | 100 | #9 (Claude Fable 5.1) |
| 2 | Claude Opus 5.5 (Max) | 100 | #3 (Claude Opus 5.5) |
| 3 | Claude Sonnet 5.5 (Max) | 100 | #11 (Claude Sonnet 5.5) |
| 4 | Claude Opus 5 (Max) | 99 | #12 (Claude Opus 5) |
| 5 | GPT-6 Astra (Max) | 99 | #6 (GPT-6 Astra) |
| 6 | GPT-6.1 Sol | 99 | #7 |
| 7 | Gemini 4 Argon | 99 | #8 |
| 8 | Claude Fable 5 (Max) | 95 | #16 (Claude Fable 5) |
| 9 | Muse Wonder Full | 94 | #10 |
| 10 | Claude Haiku 5.5 (Max) | 89 | #40 (Claude Haiku 5.5) |
| 11 | Kimi K3 | 87 | #30 |
| 12 | Muse Wonder | 84 | #15 |
| 13 | GPT-5.6 Sol (Max) | 83 | #19 (GPT-5.6 Sol) |
| 14 | GPT-6 Sol | 83 | #14 |
| 15 | Claude Sonnet 5 (Max) | 77 | #57 (Claude Sonnet 5) |
Interactive version: theaggregate.ai/benchmark?slug=vals-ai-proofbench · How It Works · Data refreshed daily, snapshot 2026-10-20.