OSWorld 2.0 300-step Partial Score
Agents · 2026-06-25
OSWorld 2.0 official leaderboard partial-credit score under a 300-step action budget, separated from binary completion accuracy.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-5.5 | 49.5 |
| Opus 4.7 | 39.8 |
| Sonnet 4.6 | 35.8 |
| MiniMax M3 | 16.6 |
| Qwen3.7-Plus | 16.6 |
| Kimi K2.6 | 14.4 |