Best LLM for math: AIME ranking
- models measured
- 54
- lead over #2
- +1.2 pts
- median
- 86.3%
Leader on AIME · read October 20
As of October 20, 2026, Google Gemini 3.1 Pro Preview leads math with 98.1% on AIME, ahead of OpenAI GPT-5.2 (96.9%) and OpenAI GPT-5.4 (96.7%). This ranking is one public board, shown as published, and is not part of the Vibe Coding Index.
| Rank | Model | AIME | Coding evidence | Coding index | Output $/Mtok |
|---|---|---|---|---|---|
| 1 | Gemini 3.1 Pro Preview Google | 98.1% | 35.3 | $12.00 | |
| 2 | GPT-5.2 OpenAI | 96.9% | 58.9 | $14.00 | |
| 3 | GPT-5.4 OpenAI | 96.7% | 37.7 | $15.00 | |
| 4 | Grok 4.20 xAI | 96.5% | Not benchmarked | $2.50 | |
| 5 | Claude Opus 4.7 Anthropic | 96.3% | 51.1 | $25.00 | |
| 6 | Claude Opus 4.6 Anthropic | 95.6% | 66.9 | $25.00 | |
| 7 | Gemini 3 Flash Preview Google | 95.6% | 75.8 | $3.00 | |
| 8 | Kimi K2.5 Moonshot AI | 95.6% | 58.5 | $2.50 | |
| 9 | GPT-5.4 Mini OpenAI | 95.6% | 28.2 | $4.50 | |
| 10 | Claude Opus 4.5 Anthropic | 95.4% | 66.8 | $25.00 | |
| 11 | Qwen3.6 Plus Qwen | 94.6% | 37.4 | $1.95 | |
| 12 | GPT-5 OpenAI | 93.4% | 72.2 | $10.00 | |
| 13 | GPT-5.1 OpenAI | 93.3% | 53.3 | $10.00 | |
| 14 | GLM 4.7 Z.ai | 93.3% | 33.6 | $1.98 | |
| 15 | GLM 4.6 Z.ai | 92.7% | 52.2 | $2.20 | |
| 16 | gpt-oss-120b OpenAI | 92.6% | 34.5 | $0.60 | |
| 17 | Qwen3.5-Flash Qwen | 92.5% | 6.2 | $0.26 | |
| 18 | Claude Sonnet 4.6 Anthropic | 92.3% | 57.8 | $15.00 | |
| 19 | GLM 5.1 Z.ai | 91.9% | 44.0 | $3.036 | |
| 20 | GLM 5 Z.ai | 91.7% | 59.7 | $1.92 | |
| 21 | GPT-5 Mini OpenAI | 91.5% | 59.8 | $2.00 | |
| 22 | MiniMax M2.7 MiniMax | 91.0% | 28.4 | $1.20 | |
| 23 | MiniMax M2.5 MiniMax | 88.8% | 59.4 | $0.95 | |
| 24 | GPT-5.4 Nano OpenAI | 88.8% | Not benchmarked | $1.25 | |
| 25 | Claude Sonnet 4.5 Anthropic | 88.2% | 59.1 | $15.00 | |
| 26 | GLM 4.5 Z.ai | 86.7% | 64.2 | $2.20 | |
| 27 | o3 Mini OpenAI | 86.5% | 52.1 | $4.40 | |
| 28 | gpt-oss-20b OpenAI | 86.0% | Not benchmarked | $0.09 | |
| 29 | 86.0% | Not benchmarked | $1.80 | ||
| 30 | Kimi K2 Thinking Moonshot AI | 85.4% | 63.4 | $2.50 | |
| 31 | o3 OpenAI | 85.3% | 70.7 | $8.00 | |
| 32 | DeepSeek V3.2 DeepSeek | 84.6% | 54.4 | $0.42 | |
| 33 | o4 Mini OpenAI | 83.7% | 68.6 | $4.40 | |
| 34 | 83.3% | 8.1 | $1.50 | ||
| 35 | Claude Haiku 4.5 Anthropic | 82.7% | 49.8 | $5.00 | |
| 36 | GPT-5 Nano OpenAI | 81.2% | 34.8 | $0.40 | |
| 37 | Claude Opus 4.1 Anthropic | 78.2% | 27.2 | $75.00 | |
| 38 | MiniMax M2.1 MiniMax | 77.9% | Not benchmarked | $1.20 | |
| 39 | o1 OpenAI | 71.5% | 61.7 | $60.00 | |
| 40 | Kimi K2 0711 Moonshot AI | 62.7% | 64.7 | $2.30 | |
| 41 | DeepSeek V3 0324 DeepSeek | 52.2% | 42.0 | $1.14 | |
| 42 | GPT-4.1 Mini OpenAI | 49.4% | 28.5 | $1.60 | |
| 43 | Mistral Large 3 2512 Mistral AI | 42.9% | Not benchmarked | $1.50 | |
| 44 | Mistral Medium 3.1 Mistral AI | 42.3% | Not benchmarked | $2.00 | |
| 45 | Gemini 2.5 Flash Lite Google | 42.1% | Not benchmarked | $0.40 | |
| 46 | GPT-4.1 OpenAI | 39.6% | 46.5 | $8.00 | |
| 47 | GPT-4.1 Nano OpenAI | 26.5% | 8.9 | $0.40 | |
| 48 | Llama 4 Maverick Meta Llama | 25.2% | 18.1 | $0.6525 | |
| 49 | Llama 4 Scout Meta Llama | 19.0% | 9.1 | $0.30 | |
| 50 | GPT-4o (2024-08-06) OpenAI | 14.0% | 23.1 | $10.00 | |
| 51 | Command A Cohere | 13.3% | Not benchmarked | $10.00 | |
| 52 | GPT-4o (2024-11-20) OpenAI | 11.9% | 18.2 | $10.00 | |
| 53 | GPT-4o-mini OpenAI | 11.5% | Not benchmarked | $0.60 | |
| 54 | Mistral Large Mistral AI | 9.2% | Not benchmarked | $6.00 |
What this measures
Can the model solve competition math problems with an exact answer?
AIME is the American Invitational Mathematics Examination, a competition for strong high school students. Each answer is a whole number, so grading is exact. The number is the share of problems solved, as run by Vals AI.
Limit: The questions and answers are public, so a model may have seen them in training. Vals AI notes that models tend to score better on the older 2024 set than on the 2025 set.
Read from AIME (Vals AI) on October 20, 2026; the board says it was last updated April 16, 2026. Each model appears once with its best published setting; rows that match no model in our catalog are left out. AIME benchmark page