AI agents went from 12% to 86% on desktop computer tasks (people: 72%), but 21% on long ones
On a standard test of everyday computer tasks, AI passed the human score in December 2025. On tasks taking people 1.6 hours, it got 20.6%.
OSWorld tests AI agents on real desktop tasks in Ubuntu apps, a browser and an office suite. The agent looks at the screen and uses the mouse and keyboard, as a person would. In the 2024 paper, people completed 72.36% of tasks and the best model 12.24%.
On the revised leaderboard (OSWorld-Verified), records climbed from 43.9% (Claude 4 Sonnet, July 2025) to 72.6% in December 2025, the first above the human figure. That entry was an agent framework (extra software around the models) using Claude Opus 4.5 and GPT-5, taking the best of 10 attempts. A single attempt reached 74.5% in February 2026. The best entry listed as a general model is Claude Fable 5 at 85.96% (August 2026).
The harder OSWorld 2.0 (June 2026) has 108 long tasks averaging 318 actions and about 1.6 hours for a person. The best reported completion rate is 20.6%.
What it means
Short, well-defined chores in common desktop apps are now within AI agents’ reach on this test. Longer jobs that take a person an hour or more are not: the best agent finishes about one in five. Handing an agent a long, multi-step computer job is still unreliable.
Caveats
- The 72% human score was measured on the original 2024 tasks, not the revised set the later records use.
- Leaderboard entries mix bare models, agent frameworks and best-of-several attempts, which aren’t directly comparable.
- Dates are leaderboard submission dates.