Benchmark

OSWorld-Verified

Measures: A multimodal agent completes 369 open-ended tasks on real Ubuntu/Windows desktops (office apps, browser, file system, multi-app workflows), checked by execution-based evaluation scripts; scored as % of tasks completed within a step budget (15/50/100 steps).

Maintainer
XLANG Lab, University of Hong Kong
Introduced
2024
Status
Superseded
Saturation
Saturated
Typical provenance
Mixed
In scoring set
No

PaperOfficial leaderboard

Leaderboard

Independent and third-party results only.

Leaderboard
Model% tasks completedProvenanceSourceEvaluated
Claude Sonnet 4.672.1%Third partyOS World Website via Epoch AI—
Kimi K2.563.3%Third partyOS World Website via Epoch AI—
Claude Sonnet 4.562.9%Third partyOS World Website via Epoch AI—
Claude Sonnet 4.558.1%Third partyOS World Website via Epoch AI—
Claude Sonnet 443.9%Third partyOS World Website via Epoch AI—
Claude Sonnet 4.542.9%Third partyOS World Website via Epoch AI—
Claude Sonnet 441.4%Third partyOS World Website via Epoch AI—
Claude 3.7 Sonnet35.8%Third partyOS World Website via Epoch AI—
Claude 3.7 Sonnet35.6%Third partyOS World Website via Epoch AI—
Computer-Using Agent (computer-use-preview)31.3%Third partyOS World Website via Epoch AI—
Claude Sonnet 431.2%Third partyOS World Website via Epoch AI—
Computer-Using Agent (computer-use-preview)30.5%Third partyOS World Website via Epoch AI—
Claude 3.7 Sonnet27.1%Third partyOS World Website via Epoch AI—
Computer-Using Agent (computer-use-preview)26.0%Third partyOS World Website via Epoch AI—
o3
OpenAI · Medium effort
23.0%Third partyOS World Website via Epoch AI—
o3
OpenAI · Medium effort
17.2%Third partyOS World Website via Epoch AI—
o3
OpenAI · Medium effort
9.1%Third partyOS World Website via Epoch AI—
Qwen2.5-VL-72B-Instruct5.0%Third partyOS World Website via Epoch AI—
Qwen2.5-VL-72B-Instruct4.4%Third partyOS World Website via Epoch AI—
  • Epoch AI, 'Capabilities & benchmarking', epoch.ai/benchmarks, CC BY 4.0.

Sources

OSWorld-Verified: leaderboard and what it measures · supermarket.dev