DynamicGUIBench (15 Steps): leaderboard
DynamicGUIBench (2026) is an online desktop GUI agent benchmark whose tasks change the interface between actions, so a single post-action screenshot can miss decisive evidence such as pop-ups, short-lived references or dynamic lists. This board is the task success rate in percent over all tasks with at most 15 interaction steps per task; higher is better.
Metric: Task success rate (%), averaged over all tasks, with at most 15 interaction steps per task, on the 149 DynamicGUIBench tasks over ten desktop applications (Chrome, GIMP, LibreOffice Calc, Impress and Writer, multi-app workflows, OS, Thunderbird, VLC, VS Code) whose interfaces change between agent actions (interruptive UI states, ephemeral references, dynamic lists, content-triggered interactions), online in a virtual machine with task-specific evaluators; higher is better. Source: arxiv.org. Saturation forecast: Around 2029. 3 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 VL 8B | 14.8 | #917 |
| 2 | Qwen 3 VL 4B | 10.7 | #1151 |
| 3 | O3 | 6.7 | #237 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=dynamicguibench-15-steps · How It Works · Data refreshed daily, snapshot 2026-11-21.