FrontierMath: leaderboard
Epoch AI's expert-crafted math benchmark featuring original, unpublished problems that test genuine mathematical reasoning rather than memorization.
Metric: Accuracy (%, 285 private v2 problems). Source: epoch.ai. Status: years away from saturation. 101 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | GPT-6 (Max) | 93.68 |
| 2 | Claude Fable 5.1 (Max) | 90.18 |
| 3 | GPT-5.6 Sol (Max) | 89.12 |
| 4 | GPT-5.5 Pro (xHigh) | 87.72 |
| 5 | Claude Fable 5 (Max) | 87.02 |
| 6 | GPT-5.6 Terra (Max) | 85.96 |
| 7 | Claude Opus 5 (Max) | 85.61 |
| 8 | GPT-5.5 (xHigh) | 85.26 |
| 9 | GPT-5.4 Pro (xHigh) | 82.46 |
| 10 | GPT-5.6 Luna (Max) | 82.11 |
| 11 | Claude Opus 4.8 (Max) | 80 |
| 12 | GPT-5.4 (xHigh) | 78.6 |
| 13 | Qwen 3.8 Max (xHigh) | 74.74 |
| 14 | Kimi K3 (Max) | 72.18 |
| 15 | Gemini 3.7 Flash (High) | 71.58 |
Interactive version: theaggregate.ai/benchmark?slug=frontiermath · How It Works · Data refreshed daily, snapshot 2026-09-05.