OSWorld-Verified: leaderboard
The July 2025 corrected release of OSWorld: 369 computer use tasks run in a real desktop virtual machine with web and desktop apps, after over 300 reported task and grading problems were fixed. An agent operates the desktop over many steps, a task specific script then inspects the machine to decide success, and scores are self reported success rates where higher is better. Source: xlang.ai/blog/osworld-verified.
Metric: Success rate (imported). Source: os-world.github.io. Source: benchmarklist.com. Saturation forecast: Around December 2027. 143 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.8 Max | 86.1 | #35 |
| 2 | Claude Fable 5 | 86 | #16 |
| 3 | Kimi K3 | 84.8 | #30 |
| 4 | Qwen 3.8 27B | 84.3 | #105 |
| 5 | Claude Opus 4.8 | 83.4 | #37 |
| 6 | Claude Opus 5 | 83.4 | #12 |
| 7 | GPT-5.6 Sol | 83 | #19 |
| 8 | Gemini 3.6 Flash | 83 | #51 |
| 9 | Claude Opus 4.7 | 82.8 | #49 |
| 10 | MiMo-V2.6-Pro | 82 | #42 |
| 11 | Claude Sonnet 5 | 81.2 | #57 |
| 12 | MiMo-V2.6-Flash | 80.8 | #70 |
| 13 | Muse Spark 1.1 | 80.8 | #52 |
| 14 | GPT-5.5 | 78.7 | #31 |
| 15 | Gemini 3.5 Flash | 78.4 | #60 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=osworld-verified · How It Works · Data refreshed daily, snapshot 2026-10-20.