ArXivMath 08/2026 (Native Harnesses)
Math · 2026-09-15
MathArena's August 2026 research-mathematics release: 57 questions, native model harnesses, Python and SageMath without internet, and Gemini 3.8 Flash answer-equivalence grading. New construction emphasizes resolved conjectures; this tool-enabled protocol is distinct from earlier monthly evaluations. Time and cost limits were introduced partway through the published runs. 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 | 96.5 |
| GPT-6 Sol | 93.0 |
| GPT-6 Astra | 88.6 |
| Claude Fable 5.1 | 87.7 |
| Claude Opus 5.5 | 86.0 |
| Gemini 4 Argon | 82.5 |
| Grok 4.7 | 46.5 |
| Muse Spark 1.3 | 45.6 |
| DeepSeek V4.1 Flash | 43.9 |
| Gemini 3.8 Flash | 40.4 |
| Kimi K3 | 38.6 |