cua-speedrun - OSWorld 2.0 (52 Tasks): leaderboard
Computer-use agents complete a 52-task subset of OSWorld 2.0 desktop tasks in identical virtual machines on Modal, with the task verifier scoring the final state. Each row is one model at one reasoning effort. Mean verifier score with partial credit, in percent; higher is better. Source: cuaspeedrun.com.
Metric: Mean verifier score (%), partial credit over the task set. Source: cuaspeedrun.com. Saturation forecast: Around February 2027. 10 models tracked. Updated 4 Nov 2026.
Top models
| # | Model | Setting | Score | Overall rank |
|---|---|---|---|---|
| 1 | Claude Opus 5.5 | Max | 85.06 | #7 |
| 2 | GPT-6 Astra | xHigh | 76.9 | #8 |
| 3 | Muse Wonder | Max | 71.28 | #31 |
| 4 | GPT-6.1 Terra | Max | 68.4 | #22 |
| 5 | Gemini 3.8 Flash | High | 60.18 | #39 |
| 6 | Claude Opus 5 | Low | 52.06 | #23 |
| 7 | Kimi K3 | Max | 47.74 | #59 |
| 8 | Claude Sonnet 5 | Low | 39.08 | #118 |
| 9 | MiniMax-M3 | Thinking | 23.19 | #278 |
| 10 | GLM-5V Turbo | Thinking | 8.44 | #362 |
No result here: #1 Claude Fable 5.2, #2 GPT-6.2 Astra, #3 Gemini 4 Carbon.
Interactive version: theaggregate.ai/benchmark?slug=cua-speedrun-osworld-2-0-52-tasks · How It Works · Data refreshed daily, snapshot 2026-11-17.