WebsiteBench - Browser: leaderboard
WebsiteBench (TIGER-AI-Lab) gives a coding agent a production-level reference website (an expert-built reproduction of a real site) that it may explore only through a browser (screenshots, clicks, typing, form submissions and the site's media, never its HTML, JavaScript, CSS, API responses or server code) and up to 16 hours to rebuild it as a full-stack application. Unlike Design2Code, which turns a screenshot into a static page, it grades the rebuild's behaviour: 70 sites carry 46,916 hidden test cases (302 to 2,016 per site), and a case passes only when its functional checks pass and every visual checkpoint reaches a normalized SSIM of 0.99. Scores pool tests rather than averaging sites: 100 x passed / total over the tasks a model's run scored. Errored or timed-out runs have no test total and drop out of that model's denominator (cohort of 17 October 2026: Kimi K3 two timeouts; DeepSeek V4.1 Flash three errors; GPT-6.1 Sol two; Claude Opus 5 and Qwen 3.8 Max one each), and Claude Opus 5.5's runs of three sites were scored on other case counts than the task list's. The leaderboard states no agent harness or reasoning effort per model. Browser group: interface responses, state changes read back later and full user workflows, including persistence, other users and recovery; some tests also compare screenshots.
Metric: Browser-group hidden tests passed, pooled (%). Source: website-bench.com. Saturation forecast: Rough model projection: around 2026. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5.5 (Benchmark Sweep Run: source=WebsiteBench; cohort=websitebench-70-20261017; source-model=claude-opus-5-5; harness=unreported; effort=unreported; case totals differ from the other rows on article, crunchyroll, everyplate) | 21.76 | |
| 2 | GPT-6.1 Sol (Benchmark Sweep Run: source=WebsiteBench; cohort=websitebench-70-20261017; source-model=gpt-6.1-sol; harness=unreported; effort=unreported) | 16.89 | |
| 3 | Claude Opus 5 (Benchmark Sweep Run: source=WebsiteBench; cohort=websitebench-70-20261017; source-model=claude-opus-5; harness=unreported; effort=unreported) | 12.87 | |
| 4 | GLM-5.3 Flash (Benchmark Sweep Run: source=WebsiteBench; cohort=websitebench-70-20261017; source-model=glm-5.3-flash; harness=unreported; effort=unreported) | 11.41 | |
| 5 | DeepSeek V4.1 Flash (Benchmark Sweep Run: source=WebsiteBench; cohort=websitebench-70-20261017; source-model=deepseek-v4.1-flash; harness=unreported; effort=unreported) | 9.8 | |
| 6 | Grok 4.6 (Benchmark Sweep Run: source=WebsiteBench; cohort=websitebench-70-20261017; source-model=grok-4.6; harness=unreported; effort=unreported) | 8.33 | |
| 7 | Kimi K3 (Benchmark Sweep Run: source=WebsiteBench; cohort=websitebench-70-20261017; source-model=kimi-k3; harness=unreported; effort=unreported) | 6.82 | |
| 8 | Qwen 3.8 Max (Benchmark Sweep Run: source=WebsiteBench; cohort=websitebench-70-20261017; source-model=qwen3.8-max; harness=unreported; effort=unreported) | 4.06 |
Interactive version: theaggregate.ai/benchmark?slug=websitebench-browser · How It Works · Data refreshed daily, snapshot 2026-11-12.