MathArena - APEX 2025: leaderboard
Twelve very hard problems collected from 2025 mathematics competitions, half from final answer contests and half rewritten from proof based ones so a single answer can be checked. They were kept only if four frontier models failed every attempt when the set was built in August 2025. Each model makes repeated attempts, only the final answer is checked, and accuracy is the share of attempts solved, higher is better.
Metric: Accuracy (%). Source: matharena.ai. Source: benchmarklist.com. Saturation forecast: Around January 2027. 48 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.8 (Max) | 81.25 | #37 (Claude Opus 4.8) |
| 2 | GPT-5.5 (xHigh) | 80.21 | #31 (GPT-5.5) |
| 3 | Hy4 preview | 74.2 | #38 |
| 4 | GPT-5.4 Pro (xHigh) | 69.79 | #32 (GPT-5.4 Pro) |
| 5 | Kimi K3 (Thinking) | 65.62 | #30 (Kimi K3) |
| 6 | Kimi K3 (Max) | 65.6 | #30 (Kimi K3) |
| 7 | DeepSeek V4.1 Flash (Max) | 65.6 | #36 (DeepSeek V4.1 Flash) |
| 8 | DeepSeek V4 Pro (0813) (Max) | 65.3 | #53 (DeepSeek V4 Pro (0813)) |
| 9 | Gemini 3.1 Pro (Preview) | 60.94 | #58 |
| 10 | DeepSeek V4 Flash (0731) (Max) | 58.6 | #78 (DeepSeek V4 Flash (0731)) |
| 11 | GPT-5.4 (xHigh) | 54.17 | #79 (GPT-5.4) |
| 12 | Qwen 3.7 Max | 44.5 | #75 |
| 13 | Claude Opus 4.7 (xHigh) | 40.62 | #49 (Claude Opus 4.7) |
| 14 | DeepSeek V4 Pro | 38.3 | #102 |
| 15 | Nex N2 Pro | 36.5 | #107 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=matharena-apex-2025 · How It Works · Data refreshed daily, snapshot 2026-10-20.