ArXivMath 09/2026 (Native Harnesses)
Math
MathArena's September 2026 research-mathematics release: 53 questions, run under the August native-harness protocol with Python and SageMath and no internet, a 12-hour and $100 limit per problem, and Gemini 3.8 Flash answer-equivalence grading. Published scores average observed correctness across problem attempts; the count records distinct problems, not a single-run binomial denominator.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6.1 Sol | 91.5 |
| Gemini 4 Argon | 87.7 |
| Claude Opus 5.5 | 82.1 |
| Gemini 3.8 Flash | 50.0 |
| Grok 4.7 | 45.3 |
| DeepSeek V4.1 Flash | 40.6 |