Mind2Web-Live: leaderboard

A model plans the actions of the WebCanvas browser agent on 104 Mind2Web tasks run on the live websites; the score is the share of tasks in which every annotated key state was reached (a three-run mean for six of the models; the paper does not give the run count for the others). Unlike Mind2Web, which compares actions to a recorded trace offline, it checks intermediate states on the live site.

Metric: Task success rate on 104 live tasks (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 13 models tracked. Updated 17 Nov 2026.

Top models

#ModelScoreOverall rank
1GPT-4 Preview (0125)23.1#1197
2GPT-4o (2024-05-13)22.1#735
3GPT-4 Turbo21.1#1008
4Claude 3 Sonnet (20240229)20.2#1983
5Qwen 1.5 110B Chat20.2#1790
6GPT-3.5 Turbo (0125)16.5#3575
7Claude 3 Haiku (20240307)16.3#2384
8Qwen 2 72B Instruct15.4#1523
9Qwen 1.5 72B Chat15.4#2220
10Claude 3 Opus (20240229)14.4#1158
11Mistral 7B Instruct (v0.3)11.7#6707
12Qwen 1.5 14B Chat10.6#3263
13Qwen 1.5 7B Chat10.6#6485

No result here: #1 Claude Fable 5.2, #2 GPT-6.2 Astra, #3 Gemini 4 Carbon.

Interactive version: theaggregate.ai/benchmark?slug=mind2web-live · How It Works · Data refreshed daily, snapshot 2026-11-17.