μ-MATH: leaderboard
μ-MATH (Toloka) — meta-evaluation companion to U-MATH testing how well models judge the correctness of math solutions; accuracy (%).
Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O1 | 89.5 | #183 |
| 2 | O1 Mini | 84.8 | #323 |
| 3 | QwQ 32B-Preview | 83.3 | #651 |
| 4 | DeepSeek R1 | 82.2 | #251 |
| 5 | Gemini 2.0 Flash (01-21) (Thinking) | 81.2 | #322 (Gemini 2.0 Flash (01-21)) |
| 6 | Gemini 1.5 Pro | 80.7 | #445 |
| 7 | GPT-4o (2024-08-06) | 77.4 | #336 |
| 8 | Mistral Large 2 (Nov) Instruct (2411) | 76.7 | #465 |
| 9 | Qwen 2.5 72B Instruct | 75.7 | #454 |
| 10 | Claude 3.5 Sonnet | 75 | #350 |
| 11 | Gemini 1.5 Flash | 74.9 | #618 |
| 12 | Qwen2.5-Math-72B-Instruct | 74.4 | #717 |
| 13 | GPT-4o Mini (2024-07-18) | 72.5 | #558 |
| 14 | Qwen 2.5 7B Instruct | 69.9 | #862 |
| 15 | Qwen2.5-Math-7B-Instruct | 63.3 | #1444 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=math · How It Works · Data refreshed daily, snapshot 2026-10-20.