MathNet: leaderboard
Metric: Problem-solving accuracy (%) over all problems (micro average) on MathNet-Solve-Test, Olympiad problems from official national competition booklets of 47 countries; a GPT-5 grader scores each solution 0-7 against the official solution and 6 or more counts as correct; image-capable models get the figures, text-only models a text description of them; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 14 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 78.4 |
| 2 | Gemini 3 Flash (Preview) | 70.4 |
| 3 | GPT-5 | 69.3 |
| 4 | GPT-5 Mini | 57 |
| 5 | Claude Opus 4.6 (Thinking) | 45.7 |
| 6 | GPT-5 Nano | 42.2 |
| 7 | Gemini 2.5 Flash | 41.1 |
| 8 | DeepSeek V3.2 (Non-reasoning) | 40.1 |
| 9 | Grok 3 | 28.5 |
| 10 | GPT-4.1 | 21.4 |
| 11 | Llama 4 Maverick Instruct FP8 | 14.7 |
| 12 | GPT-4o | 6.8 |
| 13 | Ministral 3B | 4.4 |
Interactive version: theaggregate.ai/benchmark?slug=mathnet · How It Works · Data refreshed daily, snapshot 2026-10-07.