cua-speedrun - OSWorld (50 Tasks): leaderboard
Computer-use agents complete a 50-task subset of OSWorld desktop tasks, scored by the task verifier. Each row is one model at one reasoning effort, run directly or inside Claude Code or Codex. 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: Estimated already saturated. 18 models tracked. Updated 4 Nov 2026.
Top models
| # | Model | Setting | Score | Overall rank |
|---|---|---|---|---|
| 1 | Claude Opus 5.5 | Max | 95.62 | #7 |
| 2 | GPT-6 Astra | xHigh | 91.62 | #8 |
| 3 | GPT-6.1 Terra | Max | 91.62 | #22 |
| 4 | Claude Opus 5 | High | 91.62 | #23 |
| 5 | Gemini 3.8 Flash | Low | 91.62 | #39 |
| 6 | Kimi K3 | Max | 89.62 | #59 |
| 7 | Gemini 3.7 Flash | High | 89.62 | #38 |
| 8 | Claude Sonnet 5 | High | 89.62 | #118 |
| 9 | Muse Wonder | High | 89.62 | #31 |
| 10 | GPT-5.6 Sol | xHigh | 81.62 | #33 |
| 11 | GPT-5.6 Luna | Medium | 81.62 | #130 |
| 12 | Muse Spark 1.1 | xHigh | 79.62 | #93 |
| 13 | Yutori n2 | xHigh | 79.62 | |
| 14 | Muse Spark 1.3 | xHigh | 77.62 | #45 |
| 15 | MiniMax-M3 | Non-reasoning | 75.62 | #278 |
No result here: #1 Claude Fable 5.2, #2 GPT-6.2 Astra, #3 Gemini 4 Carbon.
Example tasks
Can you make my computer bring back the last tab I shut down?
Create two pivot tables in a new sheet named "Sheet2" showing the total revenue for each product and sales channel.
Interactive version: theaggregate.ai/benchmark?slug=cua-speedrun-osworld-50-tasks · How It Works · Data refreshed daily, snapshot 2026-11-17.