DocOS - Hard: leaderboard
DocOS (Liu et al., 2026) tests GUI agents that must find the relevant page of an application’s official online documentation in a live browser and then carry out the task in the desktop application from screenshots. The board is the task completion rate on the 254 Hard tasks (the most execution steps), in percent (higher is better); a task counts when the final application state meets its goal conditions.
Metric: Task completion rate (%; share of tasks whose final application state meets the goal after the agent searches the official documentation in a live browser for up to 15 steps and then executes for up to 50 steps; the 254 Hard tasks). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | UI-TARS-1.5-7B | 13.08 | |
| 2 | MAI-UI-8B | 10 | |
| 3 | GUI-Owl-7B | 3.85 | |
| 4 | Qwen3-VL-8B | 3.85 | |
| 5 | GELab-Zero-4B-preview | 2.12 | |
| 6 | Aguvis-7B | 1.29 |
Interactive version: theaggregate.ai/benchmark?slug=docos-hard · How It Works · Data refreshed daily, snapshot 2026-11-12.