MATH Level 5 — leaderboard
The hardest tier (Level 5) of the Hendrycks MATH dataset - 1,324 competition-level problems spanning algebra, number theory, geometry, and combinatorics. Evaluated by Epoch AI.
Metric: Accuracy (%). Source: epoch.ai. Status: saturated. 108 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Human Expert | 100 |
| 2 | GPT-5 (High) | 98.13 |
| 3 | GPT-5 (Medium) | 97.92 |
| 4 | GPT-5 Mini (High) | 97.85 |
| 5 | O4 Mini (2025-04-16) (High) | 97.83 |
| 6 | O3 (2025-04-16) (High) | 97.77 |
| 7 | Claude Sonnet 4.5 (Thinking 32K) | 97.73 |
| 8 | Qwen 3 Max (2025-09-23) | 97.13 |
| 9 | GPT-5 Mini (Medium) | 96.79 |
| 10 | DeepSeek R1 0528 | 96.64 |
| 11 | O3 Mini (2025-01-31) (High) | 96.49 |
| 12 | Gemini 2.5 Pro (Preview 05-06) | 95.9 |
| 13 | Gemini 2.5 Pro (Preview 03-25) | 95.56 |
| 14 | GPT-5 Nano (Medium) | 95.24 |
| 15 | O3 Mini (Medium) | 95.17 |
Interactive version: theaggregate.ai/benchmark?slug=math-level-5 · How the rankings work · Data refreshed daily, snapshot 2026-07-22.