OSWorld 2 (Aug 8) Offline, 500-step Binary Accuracy
Agents · 2026-08-08
OSWorld 2.0 v2026.08.08 offline subset of 82 tasks runnable without internet, binary accuracy with 500 steps. Task roster is published in xlang-ai/OSWorld-V2 evaluation_examples/test_v2_offline_no_internet.json.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5 | 34.7 |
| GPT-5.6 Sol | 28.1 |