MLE-bench-30 - bronze medal pass@1
Code · 2025-08-07
OpenAI's 30-competition subset of MLE-bench (Kaggle competitions under 50GB and 10 hours, chosen for interest and diversity), run agentically with a GPU and up to 24 hours per competition. The score is the share of competitions where a single attempt earns at least a bronze medal (bronze pass@1). Higher is better.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-5.5 | 36.7 |
| GPT-5.4 | 23.3 |
| GPT-5.1-Codex-Max | 17.0 |
| GPT-5.2 | 16.0 |
| GPT-5.1 | 12.0 |
| GPT-5.2-Codex | 10.0 |
| GPT-5 | 8.0 |