ParaGUIBench: leaderboard

Metric: Success rate (%; all 233 long-horizon desktop and web tasks with parallelizable sub-tasks; single serial GUI agent (one perception-decision-action loop, no planner) on a Docker desktop with 4 GiB RAM and 2 vCPUs, each model under its official visual-history setting; verdict read from the final environment state by task-specific checks). Source: arxiv.org. Saturation forecast: Around April 2027. 4 models tracked.

Top models

#ModelScore
1Claude Sonnet 4.633.5
2Seed 1.827.9
3Qwen 3 VL 235B A22B Instruct11.2

Interactive version: theaggregate.ai/benchmark?slug=paraguibench · How It Works · Data refreshed daily, snapshot 2026-09-29.