# OSWorld 2.0

A much harder follow-up to OSWorld: the AI must carry out long, multi-step office and professional workflows on a real computer, jobs that take a person about an hour and a half. It shows whether AI can handle long, realistic computer work rather than quick clicks.

Category: agents. Type: leaderboard. Status: active.
Metric: Share of the 108 long workflows completed fully and correctly within 500 steps (binary accuracy, %) (higher is better).
Source: OSWorld 2.0 team (XLANG Lab, University of Hong Kong) — https://osworld-v2.xlang.ai/static/data/leaderboard/official-results.json, results as of 2026-11-14, retrieved 2026-11-17, license: not stated.

| Rank | Model | Organization | Reasoning effort | Tool setting | Task release | Binary accuracy (%) | Partial score (%) | Official |
|---|---|---|---|---|---|---|---|---|
| 1 | Claude Opus 5 | Anthropic | max | batch tool | v2.1 | 44.33 | 77.67 | yes |
| 2 | MuseSpark 1.3 | MuseSpark | max | batch tool | v2.1 | 39.25 | 69.56 | yes |
| 3 | Claude Opus 5 | Anthropic | high | batch tool | v2.1 | 36.89 | 68.92 | yes |
| 4 | Claude Opus 5 | Anthropic | xhigh | batch tool | v2.1 | 33.33 | 73.11 | yes |
| 5 | Claude Opus 5 | Anthropic | medium | batch tool | v2.1 | 33.01 | 64.85 | yes |
| 6 | Claude Opus 5 | Anthropic | max | batch tool | v2026.08.08 | 31.43 | 68.31 | yes |
| 7 | Claude Opus 5 | Anthropic | xhigh | batch tool | v2026.08.08 | 30.21 | 67.67 | yes |
| 8 | Claude Opus 5 | Anthropic | high | batch tool | v2026.08.08 | 28.98 | 63.79 | yes |
| 9 | GPT-5.6 Sol | OpenAI | max | batch tool | v2026.08.08 | 27.34 | 62.72 | yes |
| 10 | MiniMax M3.1 Flash | MiniMax | adaptive | batch tool | v2026.08.08 | 26.85 | 62.01 | no |
| 11 | Claude Opus 5 | Anthropic | medium | batch tool | v2026.08.08 | 25.33 | 60.49 | yes |
| 12 | Claude Opus 5 | Anthropic | low | batch tool | v2026.08.08 | 22.32 | 54.89 | yes |
| 13 | Claude Opus 4.8 | Anthropic | max | batched tool | v2026.06.24 | 20.6 | 54.8 | yes |
| 14 | Claude Opus 5 | Anthropic | low | batch tool | v2.1 | 18.81 | 59.09 | yes |
| 15 | Claude Opus 4.8 | Anthropic | max | standard | v2026.06.24 | 18.52 | 49.33 | yes |
| 16 | Claude Opus 4.7 | Anthropic | max | batched tool | v2026.06.24 | 18.2 | 48.91 | yes |
| 17 | MuseSpark 1.2 | MuseSpark | xhigh | batch tool | v2.1 | 15.84 | 44.08 | no |
| 18 | Claude Opus 4.7 | Anthropic | max | standard | v2026.06.24 | 13.9 | 49.1 | yes |
| 19 | GPT-5.5 | OpenAI | xhigh | batch tool | v2026.06.24 | 13.0 | 49.5 | yes |
| 20 | Claude Sonnet 4.6 | Anthropic | medium | standard | v2026.06.24 | 9.3 | 33.9 | yes |
| 21 | Claude Sonnet 4.6 | Anthropic | max | standard | v2026.06.24 | 8.3 | 41.5 | yes |
| 22 | MiniMax M3 | MiniMax | enabled | standard | v2026.06.24 | 4.6 | 22.3 | yes |
| 23 | Kimi 2.6 | Moonshot AI | enabled | standard | v2026.06.24 | 4.6 | 22.1 | yes |
| 24 | Qwen 3.7-Plus | Alibaba | thinking | standard | v2026.06.24 | 2.8 | 21.5 | yes |

Notes: Default leaderboard view: full 108-task set, 500-step budget, foundation end-to-end GUI models. Results span several task releases (v2026.06.24, v2026.08.08, v2.1), which are not directly comparable; the site mixes them by default. Partial score = share of checkpoints passed. Organization mapped from model family as the site does.
