Public leaderboard · Agents, tool use & computer use
OSWorld (OSWorld-Verified)
Gives an AI control of a real computer (mouse, keyboard, screen) and asks it to do everyday jobs in apps like a browser, spreadsheet or image editor. It shows how close AI is to operating a PC the way a person does.
Leaderboard
Top 3 right now
1stIntelligence-Indeed Agent90.2
2ndclaude-fable-5[1m]86.0
3rdPointer Agent w/ Opus 4.783.6
What it measures
369 tasks in real desktop applications on a virtual machine, checked by execution-based scripts, scored by success rate.
How to read the numbers
Share of the 369 desktop tasks the agent completed successfully (%), averaged over runs where several were submitted — higher is better.
Salesforce & The Australian National University & The University of Hong Kong
Agentic framework
100
63.4
2025-10-03
24
Kimi K2.5
Moonshot AI
General model
100
63.3
2026-01-31
25
claude-sonnet-4-5-20250929
Anthropic
General model
100
62.9
2025-10-31
Note: OSWorld-Verified results file behind the official leaderboard. Rows are grouped by model + max-steps setting and averaged the same way the official page does (mean, with ± population std when several runs exist), sorted by score; all approach types (general models, specialized models, agentic frameworks) included. as_of = latest result date in the file.
Source: OSWorld team (XLANG Lab, University of Hong Kong) · os-world.github.io · results as of 2026-08-01 · retrieved 2026-11-17