WebIGBench (Action Prompt): leaderboard
WebIGBench (Wu et al., 2026) asks multimodal models to write a single-file interactive webpage from screenshots of a real site (103 pages from 15 categories, 871 annotated interactions), and a browser-use UI agent then replays the reference interaction path on the generated page. Success rate in percent, counting pages where the agent reaches the final step within twice the reference number of steps; higher is better. This board uses the Action prompt (screenshots, textual instructions and the list of reference actions), temperature 0.
Metric: Success rate (%; the UI agent replays the reference interaction on the generated page and reaches the final step within twice the reference steps, 103 webpages, screenshots, textual instructions and the reference action list). Source: arxiv.org. Saturation forecast: Around February 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 69.9 | #262 |
| 2 | GPT-5 | 60.19 | #163 |
| 3 | Grok 4 | 47.57 | #306 |
| 4 | Qwen-VL-Max | 47.57 | #1226 |
| 5 | GLM-4.5V | 23.3 | #739 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=webigbench-action-prompt · How It Works · Data refreshed daily, snapshot 2026-11-12.