cua-speedrun - OSWorld 2.0 (52 Tasks): leaderboard
Metric: Task success (%), one run per task in the cua-speedrun Modal harness. Source: cuaspeedrun.com. Saturation forecast: Around January 2027. 21 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | GPT-6 Astra (xHigh) | 76.9 |
| 2 | GPT-6 Astra (High) | 75.02 |
| 3 | GPT-6 Astra (Low) | 68.17 |
| 4 | GPT-6 Astra (Medium) | 67.18 |
| 5 | Gemini 3.8 Flash (High) | 60.18 |
| 6 | Gemini 3.8 Flash (Medium) | 59.93 |
| 7 | Claude Opus 5 (Low) | 52.06 |
| 8 | Claude Sonnet 5 (Low) | 39.08 |
| 9 | Gemini 3.8 Flash (Low) | 35.29 |
| 10 | Kimi K3 (Max) | 28.51 |
| 11 | Kimi K3 (High) | 22.6 |
| 12 | Kimi K3 (Low) | 19.19 |
| 13 | MiniMax M3 (Thinking) | 4.1 |
Interactive version: theaggregate.ai/benchmark?slug=cua-speedrun-osworld-2-0-52-tasks · How It Works · Data refreshed daily, snapshot 2026-10-03.