Image: TIGER-AI-Lab
Why it mattersA product that promises to clone a website with AI is a common pitch. The best model today still rebuilds only one of four sites correctly.
TIGER-AI-Lab has released WebsiteBench, a test for AI agents that rebuild websites. The lab says it is the first benchmark that scores both what a site looks like and how it behaves. The repository has 190 stars and is under an Apache 2.0 license. The dataset is on Hugging Face.
The benchmark uses 70 real websites. Each one carries hidden test cases, from 302 up to 2,016 per site. In total there are 46,916 test cases. A case passes only when the functional checks work. Every visual checkpoint also has to match at a normalized SSIM of 0.99 or better.
Who passes and who does not
The published leaderboard runs eight models with a 16-hour time limit per site. Claude Opus 5.5 leads at a 26.2 percent pass rate, according to the leaderboard. GPT 6.1 Sol is next at 20.2 percent. Claude Opus 5 sits at 17.1 percent. The rest of the list includes GLM 5.3 Flash, DeepSeek V4.1 Flash, Grok 4.6, Kimi K3 and Qwen 3.8 Max.
The leaderboard also reports the cost of each run. TIGER-AI-Lab says Claude Opus 5.5 spent 305 dollars on average per site. It used 1.1 billion tokens and took 2,235 browser turns. GPT 6.1 Sol spent 35.8 dollars and took 1,023 turns. The cheap open models came in near five dollars, but their pass rates were lower.
What the pass criterion asks
The benchmark checks the rebuilt site against the original on two axes. It runs the functional tests the original site passes. It also measures visual match with SSIM, and the pass threshold is 0.99. A page with a small text shift or a missing hover state drops below that.
The authors' framing in the repository is that this is a reverse-engineering task. An agent reads the live site through a browser and then writes a full-stack rebuild from what it saw. The 16-hour cap is the compute budget. Nothing in the benchmark lets the agent read the original source.
What a 26 percent score means for a team
A product that promises to clone a site with AI is a common sales line in 2026. The best frontier model in this test rebuilds only one in four sites to a tight visual target. The number comes from a research lab's own benchmark. A buyer of such a product should ask what the vendor measures against. A side-by-side screenshot is a weaker test than SSIM 0.99. A team paying for a web-rebuild agent should run its own site against the agent before signing.
Source
- TIGER-AI-Lab/WebsiteBench, TIGER-AI-Lab
- WebsiteBench leaderboard, TIGER-AI-Lab
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
