WebsiteBenchWebsiteBench

WebsiteBenchWebsiteBench

Can AI agents rebuild websites through browser exploration?

Given a production-level website they can explore through a browser, agents must reverse-engineer a full-stack version that matches its appearance and behavior as closely as possible.

Leaderboard

WebsiteBench benchmark leaderboard: hidden-test pass rates over the 70 tasks and mean resource use per task for eight models.
# Model
1Claude Opus 5.5Anthropic, proprietary model26.271.221.835.4$305.11,095.6M2,234.8
2GPT 6.1 SolOpenAI, proprietary model20.264.316.936.8$35.8136.5M1,023.1
3Claude Opus 5Anthropic, proprietary model17.158.812.928.8$202.0352.6M807.8
4GLM 5.3 FlashZ.ai, open-source model14.743.311.424.8$4.9154.2M565.2
5DeepSeek V4.1 FlashDeepSeek, open-source model13.539.99.829.7$4.6133.1M583.3
6Grok 4.6xAI, proprietary model12.139.38.329.5$75.037.1M396.5
7Kimi K3Moonshot AI, open-source model10.641.06.825.6$36.183.7M574.1
8Qwen 3.8 MaxQwen Team, open-source model7.328.74.121.8$9.630.3M253.5

Citation

@misc{zhang2026websitebench,
  title  = {WebsiteBench: Can AI Agents Rebuild Websites through Browser Exploration?},
  author = {{WebsiteBench Authors}},
  year   = {2026},
  url    = {https://github.com/TIGER-AI-Lab/WebsiteBench},
}