MATH-Perturb (Hard): leaderboard
Perturbed versions of MATH competition problems testing genuine mathematical reasoning vs. pattern matching on memorized solutions.
Metric: Accuracy (%). Source: math-perturb.github.io. Saturation forecast: Estimated already saturated. 37 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro (Preview) | 88.89 | #130 |
| 2 | Grok 4 | 87.1 | #181 |
| 3 | Grok 3 Mini (High) | 87.1 | #281 (Grok 3 Mini) |
| 4 | O4 Mini (Medium) | 87.1 | #180 (O4 Mini) |
| 5 | QwQ-32B | 86.02 | #421 |
| 6 | DeepSeek R1 | 85.19 | #251 |
| 7 | O3 Mini (High) | 84.35 | #272 (O3 Mini) |
| 8 | O3 Mini (Medium) | 82.92 | #272 (O3 Mini) |
| 9 | DeepSeek R1 Distill Qwen 14B | 81.84 | #844 |
| 10 | O1 Mini | 79.69 | #323 |
| 11 | DeepSeek R1 Distill Llama 70B | 79.57 | #594 |
| 12 | Grok 3 | 79.21 | #304 |
| 13 | O3 Mini (Low) | 78.49 | #272 (O3 Mini) |
| 14 | O1 | 77.18 | #183 |
| 15 | Claude 3.7 Sonnet | 76.11 | #249 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=math-perturb-hard · How It Works · Data refreshed daily, snapshot 2026-10-20.