ProofBench: leaderboard
200 Lean 4 proof problems (100 public, 100 private) from graduate qualifying exams and textbooks in analysis, algebra, number theory and logic; Vals AI, 2026; pass only if it compiles without sorry.
Metric: Accuracy (%). Source: www.vals.ai. Status: saturation imminent. 31 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude Fable 5.1 (Max) | 100 |
| 2 | Claude Opus 5 (Max) | 99 |
| 3 | GPT-6 (Max) | 99 |
| 4 | Claude Fable 5 (Max) | 95 |
| 5 | Kimi K3 | 87 |
| 6 | GPT-5.6 Sol (Max) | 83 |
| 7 | Claude Sonnet 5 (Max) | 77 |
| 8 | GPT-5.6 Terra (xHigh) | 74 |
| 9 | GPT-5.6 Terra (Max) | 74 |
| 10 | GPT-5.6 Luna (Max) | 60 |
| 11 | Qwen 3.8 Max | 58 |
| 12 | Gemini 3.7 Flash (High) | 58 |
| 13 | DeepSeek V4 Flash (0731) | 56 |
| 14 | Grok 4.6 (High) | 51 |
| 15 | DeepSeek V4 Pro (0813) (Max) | 50 |
Interactive version: theaggregate.ai/benchmark?slug=proofbench · How It Works · Data refreshed daily, snapshot 2026-09-05.