All benchmarksComputer use In plain words Example taskSource
OSWorld-Verified
Get everyday jobs done on a real desktop computer.
- Tasks
- 369
- Released
- Jul 2025
- Made by
- The University of Hong Kong, Salesforce Research, Carnegie Mellon University and University of Waterloo
- Score
- Percent of tasks completed
Like asking an assistant to use your computer: open apps, click, type, and finish a chore such as updating a spreadsheet.
Update a bookkeeping spreadsheet with recent expenses taken from receipt images in a folder.
A plain-language request on a real computer with apps already set up. The AI sees the screen and clicks or types.How it is marked
A custom script for each task checks the computer's final state, such as the contents of a saved file.
Humans: People completed about 72% of tasks (measured on the original OSWorld) source
Why it matters
It measures whether AI can handle real office and computer work, not just chat about it.
Good to know
- Verified is a cleaned-up upgrade of the 2024 OSWorld; the human score comes from the original version.
- Eight Google Drive tasks may need manual setup, so some results use 361 tasks.
- OSWorld 2.0 has since been released.