OSWorld-VerifiedAgents

OSWorld-Verified is the first scalable benchmark based on real computer environments. It evaluates multimodal agents on 369 open-ended tasks across Ubuntu, Windows, and macOS, spanning web apps, desktop software, file operations, and cross-application workflows. The human completion rate is 72.36%.

Ranking Sources

#ModelScore

1

Qwen3.8-MaxAlibaba

86.1%

2

Claude Mythos PreviewAnthropic

85.4%

3

Claude Fable 5Anthropic

85.0%

4

Kimi K3Moonshot AI

84.8%

5

Qwen3.8-27BAlibaba

84.3%

6

Claude Opus 4.8Anthropic

83.4%

7

GPT-5.6 SolOpenAI

83.2%

8

Gemini 3.6 FlashGoogle

83.0%

9

Claude Opus 4.7Anthropic

82.8%

10

Claude Sonnet 5Anthropic

81.2%

36 models total

Back to benchmarks