WebGym: leaderboard
Visual web agents solving realistic tasks on live websites from screenshots, on the 1,167 held-out WebGym tasks drawn from websites absent from its training set (Bai et al., 2026). Task success rate in percent, where a GPT-4o rubric judge credits a trajectory only when every criterion is met; agents act with screen coordinates under the memory prompt; higher is better.
Metric: Task success rate (%; 1,167 held-out tasks, each from a website absent from the training set, on live websites; memory prompt with coordinate-grounded actions; success only when a GPT-4o rubric judge finds every criterion met). Source: arxiv.org. Saturation forecast: Around February 2027. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 VL 8B (Thinking) | 28.2 | #887 |
| 2 | Qwen 3 VL 8B Instruct | 26.2 | #982 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=webgym · How It Works · Data refreshed daily, snapshot 2026-11-12.