ComponentBench (Pixel): leaderboard

Metric: Pass rate (%; share of tasks whose programmatic verifier passes; screenshot only, actions by predicted coordinates, in the shared BrowserGym harness or, for UI-TARS-1.5-7B, its native pixel-only harness; ComponentBench-Full: 2,910 programmatically verified tasks on 97 canonical UI components across widely used web component libraries; 20-step budget; one deterministic run). Source: arxiv.org. Saturation forecast: Around December 2026. 7 models tracked.

Top models

#ModelScore
1Gemini 3 Flash85.4
2GPT-5.483.8
3GPT-5.4 Mini77.1
4Gemini 3.1 Flash Lite63.3
5Qwen 3 VL 235B A22B50.5
6GPT-5 Mini48.9
7uitars-1.5-7B12.6

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