OSWorld 2.0 is a benchmark from XLANG Lab that evaluates computer-use agents on 108 long-horizon tasks in real operating-system environments, spanning desktop and web applications across professional domains. Its tasks are far longer than the original OSWorld — a skilled human takes a median of roughly 1.6 hours per task — and are graded against many weighted checkpoints rather than a single pass/fail check. It is a separate, harder benchmark from the original OSWorld.

We source results from the public OSWorld 2.0 leaderboard. Our chart reports binary accuracy: the percentage of tasks an agent completes in full, where every scoring checkpoint for that task passes. The leaderboard also reports a partial score (the average fraction of checkpoints satisfied per task), which we keep alongside binary accuracy in the data export.

OSWorld 2.0 runs agents in real OS environments and grades long-horizon workflows against an average of about 27 weighted checkpoints per task. We report results at the leaderboard’s default budget of 500 agent steps. Because the tasks are long and binary accuracy requires finishing the entire workflow, scores are low even for frontier models. We also keep each run’s reasoning setting, tool setting, step budget, and estimated cost in the data export.