OSWorld 2.0

Operating real desktop applications from screenshots to complete multi-step tasks.

CategoryAgents
MetricAccuracy
SourceEpoch AI (evaluator)
LicenseCC BY 4.0
Benchmark pageosworld-v2.xlang.ai
StatusScored
Weight1.0 within Agents
FitDiscrimination 1.58, noise 0.25

Results

#ModelProviderResultDate
1 Claude Opus 5 Max Anthropic 31.4% 2026-11-11
2 Claude Opus 5 Extra High Anthropic 30.2% 2026-11-11
3 Claude Opus 5 High Anthropic 29.0% 2026-11-11
4 GPT-5.6 Sol Max OpenAI 27.3% 2026-11-11
5 Claude Opus 5 Medium Anthropic 25.3% 2026-11-11
6 Claude Opus 5 Low Anthropic 22.3% 2026-11-11
7 Claude Opus 4.8 Max Anthropic 20.6% 2026-11-11
8 Claude Opus 4.7 Max Anthropic 18.2% 2026-11-11
9 GPT-5.5 Extra High OpenAI 13.0% 2026-11-11
10 Claude Sonnet 4.6 Medium Anthropic 9.3% 2026-11-11
11 Claude Sonnet 4.6 Max Anthropic 8.3% 2026-11-11
12 Kimi K2.6 Moonshot AI 4.6% 2026-11-11
13 MiniMax-M3 MiniMax 4.6% 2026-11-11
14 Qwen3.7 Plus Alibaba 2.8% 2026-11-11