SPACEXAI
4/5 workers samplingPOSSIBLE_MODEL_CHANGE · gpt-5
public · agents

OSWorld

Desktop computer-use tasks across apps and file systems.

Items
369
Revision
2026.03
Contamination
watched
Scorer
scorer.osworld.v4
Dataset revision
[email protected]
Scorer
desktop-oracle · scorer.osworld.v4
License
MIT
Kind
Public reproduction suite
ModelScore$/taskMedian sToolsEvidence
Claude 4.1 Opus54.8$0.32849.83.8LIVE_MEASURED
Claude 4.1 Sonnet48.2$0.05833.63.4LIVE_MEASURED
Grok 4 Heavy47.6$0.15064.52.8LIVE_MEASURED
GPT-546.8$0.08137.13.0LIVE_MEASURED
Grok 444.2$0.04829.63.2LIVE_MEASURED
Gemini 2.5 Pro43.6$0.02924.92.9LIVE_MEASURED
o4-pro38.2$0.695104.92.2LIVE_MEASURED
Mistral Large 336.8$0.02228.42.9LIVE_MEASURED
DeepSeek V3.133.6$0.005122.82.7REPRODUCED
Qwen3-Max32.1$0.007027.02.8LIVE_MEASURED
Llama 4 405B31.4$0.007025.92.6REPRODUCED
GPT-5 Mini28.1$0.004414.32.4LIVE_MEASURED
Gemini 2.5 Flash27.8$0.00169.42.2LIVE_MEASURED
Grok 3 Mini22.4$0.002511.12.1LIVE_MEASURED
Llama 4 70B21.2$0.00165.32.0REPRODUCED