OSWorld 2.0
Real desktop apps across Ubuntu, Windows and macOS, driven from screenshots.
Agent Index weight
6%
in the Computer use pillar
Models tested
9
from the benchmark’s leaderboard and reports
Results
| # | Model | Score | Normalised | Harness | Cost/task | Date |
|---|---|---|---|---|---|---|
| 1 | Claude Opus 5 Supported Anthropic | 31.4% | source | — | — | |
| 2 | GPT-5.6 Sol Supported OpenAI | 27.3% | source | — | — | |
| 3 | Claude Opus 4.8 Supported Anthropic | 20.6% | source | — | — | |
| 4 | Claude Opus 4.7 Supported Anthropic | 18.2% | source | — | — | |
| 5 | GPT-5.5 Supported OpenAI | 13.0% | source | — | — | |
| 6 | Claude Sonnet 4.6 Supported Anthropic | 9.3% | source | — | — | |
| 7 | Kimi K2.6 Estimated Moonshot AI | 4.6% | source | — | — | |
| 8 | MiniMax-M3 Estimated MiniMax | 4.6% | source | — | — | |
| 9 | Qwen3.7 Plus Alibaba (Qwen) | 2.8% | source | — | — |
For models without a OSWorld 2.0 result, the index estimates one from Epoch’s Capabilities Index using a logistic fit on 9 models (residual SD 0.3 on the logit scale). Estimates are labelled and widen the model’s range.