ArXivMath 06/2026
Math
This MathArena track evaluates models on 49 research-level mathematics problems sourced from arXiv papers submitted in June 2026. Published scores average correctness across observed problem attempts; the item count records distinct problems, not a single-run binomial denominator.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6.1 Sol | 94.4 |
| GPT-6 Astra | 94.4 |
| Claude Opus 5 | 91.3 |
| Claude Fable 5.1 | 91.0 |
| GPT-6 Sol | 91.0 |
| GPT-5.6 Sol | 88.5 |
| Claude Fable 5 | 85.4 |
| GPT-5.5 | 83.6 |
| Claude Opus 5.5 | 83.3 |
| Muse Spark 1.3 | 76.7 |
| Kimi K3 | 71.5 |
| Grok 4.7 | 70.1 |
| Opus 4.8 | 70.0 |
| Gemini 3.8 Flash | 67.4 |
| Gemini 3.1 Pro Preview | 66.7 |
| Grok 4.5 | 66.7 |
| Mistral Large 4 Preview | 66.0 |
| Gemini 3.7 Flash | 62.5 |
| DeepSeek V4.1 Flash | 57.6 |
| Gemini 3.6 Flash | 56.9 |
| Muse Spark 1.1 | 56.9 |
| Gemini 3.5 Flash | 52.1 |
| GLM-5.2 | 46.5 |
| DeepSeek-V4-Flash | 43.8 |
| Step 3.7 Flash | 30.6 |
| Qwen3.6 35B A3B | 27.1 |
| Qwen3.5-2B | 4.2 |