FormalProofBench: leaderboard
Metric: Accuracy (%): share of FormalProofBench's 200 private graduate-level problems for which the model, in an agent harness with Mathlib search and Lean execution tools and up to 40 turns, submits a Lean 4 proof the kernel accepts (no new axioms); provider default sampling; higher is better. Source: arxiv.org. Saturation forecast: Around January 2028. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.5 (Thinking) | 33.5 | #79 (Claude Opus 4.5) |
| 2 | Gemini 3 Pro (Preview) | 18.5 | #64 |
| 3 | Claude Sonnet 4.5 (Thinking) | 18 | #138 (Claude Sonnet 4.5) |
| 4 | GPT-5 | 17 | #91 |
| 5 | Gemini 3 Flash (Preview) | 15 | #78 |
| 6 | GPT-5.2 | 13.5 | #105 |
| 7 | GPT-5 Nano | 12.5 | #415 |
| 8 | GPT-5.1 Codex Max | 12 | #94 |
| 9 | DeepSeek V3.2 (Non-reasoning) | 8 | #198 (DeepSeek V3.2) |
| 10 | GPT-5 Mini | 7 | #176 |
| 11 | GLM-4.7 | 4.5 | #185 |
| 12 | Grok 4.1 Fast (Reasoning) | 4 | #208 (Grok 4.1 Fast) |
| 13 | DeepSeek V3.2 (Thinking) | 3 | #198 (DeepSeek V3.2) |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=formalproofbench · How It Works · Data refreshed daily, snapshot 2026-10-11.