AIME 2025 (MathArena repeated-run mean)
Math · 2025-02-12
MathArena's AIME 2025 protocol reports the mean correctness across repeated runs per question, normally four attempts. The 30-item count records unique questions, not independent single-run Bernoulli trials.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-5.2 | 100.0 |
| Step 3.5 Flash | 98.3 |
| Gemini 3 Flash Preview | 97.5 |
| GLM-5 | 96.7 |
| DeepSeek-V3.2-Speciale | 95.8 |
| Kimi K2.5 | 95.8 |
| Gemini 3 Pro Preview | 95.0 |
| GPT-5 | 95.0 |
| DeepSeek-V3.2 | 94.2 |
| GPT-5.1 | 94.2 |
| GLM-4.5 | 93.3 |
| Grok 4 | 92.5 |
| Kimi K2 Thinking | 92.5 |
| DeepSeek-V3.2-Exp | 91.7 |
| GLM 4.6 | 91.7 |
| O4 Mini | 91.7 |
| DeepSeek-V3.1 | 90.8 |
| Grok 4 Fast | 90.8 |
| gpt-oss-120b | 90.0 |
| DeepSeek-R1-0528 | 89.2 |
| gpt-oss-20b | 89.2 |
| Grok 4.1 Fast | 89.2 |
| o3 | 89.2 |
| Gemini 2.5 Pro | 88.3 |
| GPT-5 Mini | 87.5 |
| Falcon H1R 7B | 86.7 |
| o3-mini | 86.7 |
| GPT-5 Nano | 85.0 |
| Sonnet 4.5 | 84.2 |
| Gemini 2.5 Pro Preview 05-06 | 83.3 |
| GLM 4.5 Air | 83.3 |
| K2-Think | 83.3 |
| Grok 3 Mini | 81.7 |
| o1 (2024-12-17) | 81.7 |
| Qwen3-235B-A22B | 80.8 |
| QED-Nano | 77.5 |
| Gemini 2.5 Flash | 70.8 |
| Opus 4 | 70.0 |
| DeepSeek-R1 | 70.0 |
| Qwen3-30B-A3B | 70.0 |