Atlas

Benchmarks

← All benchmarks

OSWorld 2.0 300-step Binary Accuracy

Agents · 2026-06-25

OSWorld 2.0 official leaderboard binary completion accuracy under a 300-step action budget over the 108-task v2 suite.

Top models (higher is better)

ModelScore
Opus 4.713.0
GPT-5.513.0
Sonnet 4.68.3
Kimi K2.64.6
MiniMax M33.7
Qwen3.7-Plus1.9
Loading Atlas data…