MathArena - IMProofBench Final Answers: leaderboard
Final-answer accuracy on 39 research-level IMProofBench problems with definite numerical or symbolic answers, enabling automated grading of mathematical reasoning.
Metric: Accuracy (%). Source: matharena.ai. Saturation forecast: Around January 2027. 17 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 83.52 | #58 |
| 2 | Claude Opus 4.6 (High) | 80.74 | #63 (Claude Opus 4.6) |
| 3 | GPT-5.2 (xHigh) | 73.17 | #110 (GPT-5.2) |
| 4 | Gemini 3 Pro (Preview) | 72.28 | #66 |
| 5 | GPT-5.1 (High) | 62.01 | #135 (GPT-5.1) |
| 6 | GPT-5 Pro | 59.29 | #81 |
| 7 | Grok 4 | 57.48 | #181 |
| 8 | GPT-5 (High) | 53.8 | #98 (GPT-5) |
| 9 | O4 Mini (High) | 45.22 | #180 (O4 Mini) |
| 10 | Grok 4 Fast (Reasoning) | 44.05 | #254 (Grok 4 Fast) |
| 11 | O3 (High) | 41.27 | #124 (O3) |
| 12 | Gemini 2.5 Pro | 39.15 | #152 |
| 13 | Claude Opus 4.1 (Thinking) | 38.42 | #140 (Claude Opus 4.1) |
| 14 | Gemini 2.5 Flash (Thinking) | 37.46 | #245 (Gemini 2.5 Flash) |
| 15 | GPT-OSS-120B (High) | 37.32 | #341 (GPT-OSS-120B) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=matharena-improofbench-final-answers · How It Works · Data refreshed daily, snapshot 2026-10-20.