| Rank | Model | Provider | Score |
|---|---|---|---|
| 1 | o1 preview | OpenAI | 94.8% |
| 2 | o1 mini | OpenAI | 90.0% |
| 3 | Qwen2.5 72B | Alibaba | 83.1% |
| 4 | GPT 4o | OpenAI | 76.6% |
| 5 | Llama 3.1 405B | Meta | 73.8% |
| 6 | DeepSeek V2.5 | DeepSeek | 72.7% |
| 7 | Claude 3.5 Sonnet | Anthropic | 71.1% |
| 8 | GPT 4o mini | OpenAI | 70.2% |
| 9 | Mistral Large 2 | Mistral AI | 68.6% |
| 10 | Llama 3.1 70B | Meta | 67.9% |
| 11 | Gemini 1.5 Pro | 67.7% | |
| 12 | Claude 3 Opus | Anthropic | 60.1% |
| 13 | Gemini 1.5 Flash | 54.9% | |
| 14 | Llama 3.1 8B | Meta | 51.9% |
Figures are taken from public model cards, technical reports and vendor announcements, and are approximate where sources disagree. Results from different benchmark versions or tool-use settings never share a column: Terminal Bench 2.1, 3.0 and DeepSWE v1.1 are distinct benchmarks, and HLE with tools is separate from HLE without tools. Only models with a published result are listed; the score column is shaded relative to the best and worst result on this benchmark.