WebsiteBench
Explore score history ↗Agents rebuild production websites after exploring them in a browser. The leaderboard reports hidden-test pass rates across 70 tasks. Each task has 302 to 2,016 cases; a case passes only when its functional checks pass and every visual checkpoint reaches normalized SSIM 0.99. Overall pools passed and total tests across tasks rather than averaging task rewards. The table does not state model run dates, harnesses or execution budgets.
- Evaluates
- coding agent
- Openness
- unknown
- Importer
- Model reports
WebsiteBench links
WebsiteBench results and reported scores
Reported benchmark scores · Model reports
Reported scores: 8
| Model | Organization | Reported value | Reported | Evidence |
|---|---|---|---|---|
| Claude Opus 5 | Anthropic | 17.1 | 2026-10-17 | evidence |
| Claude Opus 5.5 | Anthropic | 26.2 | 2026-10-17 | evidence |
| DeepSeek-V4.1-Flash | DeepSeek | 13.5 | 2026-10-17 | evidence |
| GLM-5.3-Flash | Z.ai | 14.7 | 2026-10-17 | evidence |
| GPT-6.1 Sol | OpenAI | 20.2 | 2026-10-17 | evidence |
| Grok 4.6 | xAI | 12.1 | 2026-10-17 | evidence |
| Kimi K3 | Moonshot AI | 10.6 | 2026-10-17 | evidence |
| Qwen 3.8 Max | Qwen | 7.3 | 2026-10-17 | evidence |
Scores are partitioned by the source that reported them and are never merged into a single cross-source ranking, because the sources measure different things and say so.
Which sources cite WebsiteBench?
- WebsiteBench official leaderboard (2026-10-17 cohort) TIGER-AI-Lab, 2026-10-17