VG-GUI-Bench (No Video): leaderboard
Video-guided mobile GUI agents on VG-GUI-Bench (Fan, Liu, Yin, Guo and Yang, ECCV 2026): for each of the 1,071 steps of 100 tutorial-paired episodes built on MONDAY, the model predicts the next tap, swipe, text entry, key press, zoom or finish from the task, its action history and the current screen. It sees no tutorial frames, only the task and the screen. Step accuracy in percent gives 0.3 for the right action type and up to 0.7 more for its arguments; higher is better.
Metric: Step accuracy (%; 0.3 for the right action type plus up to 0.7 for its arguments, over the 1,071 steps of 100 mobile GUI episodes; no tutorial frames, current screen only). Source: arxiv.org. Saturation forecast: Around February 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 58.51 | #91 |
| 2 | Kimi K2.5 | 57.22 | #252 |
| 3 | GPT-5 Mini | 55.76 | #349 |
| 4 | Claude Sonnet 4.6 | 44.35 | #158 |
| 5 | Seed 2.0 Pro | 35.93 | #209 |
| 6 | Qwen 3 VL 235B A22B | 25.9 | #477 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=vg-gui-bench-no-video · How It Works · Data refreshed daily, snapshot 2026-11-12.