BrokenMath - Utility: leaderboard
The same models solve the original, unaltered competition problems behind BrokenMath; the number is the share solved (proofs graded by OPC-R1-8B). It is the capability control reported beside the sycophancy rate.
Metric: Utility (%; accuracy on the original unperturbed problems, higher is better). Source: arxiv.org. Saturation forecast: Around June 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 (High) | 58.2 | #167 (GPT-5) |
| 2 | Grok 4 Fast (Reasoning) | 51.6 | #441 (Grok 4 Fast) |
| 3 | DeepSeek V3.1 (Thinking) | 48.4 | #497 (DeepSeek V3.1) |
| 4 | Gemini 2.5 Pro | 48.2 | #266 |
| 5 | GPT-OSS-120B (High) | 47.4 | #700 (GPT-OSS-120B) |
| 6 | Grok 4 | 46.8 | #307 |
| 7 | O4 Mini (High) | 43.8 | #353 (O4 Mini) |
| 8 | Qwen 3 4B 2507 (Thinking) | 33.5 | #1390 |
| 9 | DeepSeek R1 0528 Qwen3 8B | 32.3 | #2118 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=brokenmath-utility · How It Works · Data refreshed daily, snapshot 2026-11-13.