CAP (Cross-Site Browser Tasks): leaderboard

CAP (Macau University of Science and Technology, FellouAI and others, 2026) evaluates browser agents on realistic cross-site workflows over live websites that require complex UI actions and visual perception of dynamically rendered content, scored here on its fixed 198-task evaluation set. Each task becomes a rubric tree of execution and perception checkpoints scored by a GPT-4o judge agent with URL-grounded verification; partial completion is the mean rubric score in percent, higher is better.

Metric: Partial completion (%; mean root-node score of the rubric tree, where a failed critical child zeroes its parent and other children are averaged; a fixed 198-task evaluation set of CAP cross-site browser tasks on live websites (the benchmark holds 420 tasks over 108 websites in 24 domains); each task becomes a hierarchical rubric tree of execution, perception and format checkpoints, scored by a GPT-4o judge agent with URL-grounded verification; Browser-Use agent framework with default configuration and at most 50 reasoning-action steps). Source: arxiv.org. Saturation forecast: Around February 2028. 3 models tracked.

Top models

No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.

Interactive version: theaggregate.ai/benchmark?slug=cap-cross-site-browser-tasks · How It Works · Data refreshed daily, snapshot 2026-11-13.