Benchmarks · Coding · Agents & computer use
WebsiteBench
Rebuilding a production website from browser exploration alone: the agent sees screenshots and can click, type and navigate, but never the site's HTML, JavaScript, CSS or server code, and must write a full-stack copy that looks and behaves the same. 70 expert-built sites checked by hidden request, browser and system tests; share of tests passed, pooled over all tasks, with mean API cost per task. From TIGER-AI-Lab.
8 models · 8 results · data as of 2026-11-13 · website-bench.com ↗
Leaderboard
Heavy line: best score at each price. Thin lines join one model's effort levels. Cost uses each model's median subscription plan, or API price with no plan data. 4 more models under All models in settings.
More from this benchmark
7 published measuresScore vs any published measure7 measures: sub-scores, splits, counts
5 more models under All models in settings.
Breakdown
darker is better within each column · click a column to sort| Model | ||||||
|---|---|---|---|---|---|---|
| Opus 5.5 | 71.2 | 21.8 | 35.4 | 2,235 | 11,812 | 45,091 |
| GPT-6.1 Sol | 64.3 | 16.9 | 36.8 | 1,023 | 8,870 | 43,970 |
| Opus 5 | 58.8 | 12.9 | 28.8 | 807.8 | 7,930 | 46,379 |
| GLM 5.3 Flash | 43.3 | 11.4 | 24.8 | 565.2 | 6,875 | 46,916 |
| DeepSeek V4.1 Flash | 39.9 | 9.8 | 29.7 | 583.3 | 5,990 | 44,485 |
| Grok 4.6 | 39.3 | 8.33 | 29.5 | 396.5 | 5,699 | 46,916 |
| Kimi K3 | 41.0 | 6.82 | 25.6 | 574.1 | 4,824 | 45,345 |
| Qwen 3.8 Max | 28.7 | 4.06 | 21.8 | 253.5 | 3,372 | 46,216 |
All results
8 rows| # | Model | Effort | WebsiteBench | Subscription $ / task | API $ / task |
|---|---|---|---|---|---|
| 1 | Opus 5.5 | 26.2% | $4.62 | $305 | |
| 2 | GPT-6.1 Sol | 20.2% | $1.20 | $36 | |
| 3 | Opus 5 | 17.1% | $3.50 | $202 | |
| 4 | GLM 5.3 Flash | 14.7% | free | $4.88 | |
| 5 | DeepSeek V4.1 Flash | 13.5% | $0.33 | $4.62 | |
| 6 | Grok 4.6 | 12.1% | $7.71 | $75 | |
| 7 | Kimi K3 | 10.6% | $3.28 | $36 | |
| 8 | Qwen 3.8 Max | 7.3% | $1.14 | $9.63 |