WindowsWorld (Screenshot + Accessibility Tree): leaderboard
Metric: Final task success rate (%) over all 181 tasks of WindowsWorld (persona-driven Windows desktop tasks across 17 applications, executed with a unified PyAutoGUI action space and level-specific step budgets of 15, 25, 40 and 20), the model observing screenshots plus the accessibility tree, states judged by a human-validated Qwen3-VL-Plus judge; infeasible Level 4 tasks count as solved when the agent correctly rejects them; higher is better. Source: arxiv.org. Saturation forecast: Around January 2028. 5 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3 Flash (Preview) | 20.44 |
| 2 | Gemini 3 Pro (Preview) | 14.77 |
| 3 | GPT-5.2 | 1.12 |
| 4 | Claude Sonnet 4.5 | 0.55 |
Interactive version: theaggregate.ai/benchmark?slug=windowsworld-screenshot-plus-accessibility-tree · How It Works · Data refreshed daily, snapshot 2026-10-07.