Reported News
WebsiteBench: best AI agent passes 26.2% of hidden tests rebuilding 70 real websites
WebsiteBench, from TIGER-AI-Lab, asks AI agents to rebuild 70 production websites as full-stack apps just by exploring them in a browser, scored on 46,916 hidden test cases. On its leaderboard the top model, Claude Opus 5.5, passes 26.2% overall at a mean $305.1 per task; GPT 6.1 Sol scores 20.2% at $35.8.
- Changes reported
- 6
- listed below, each from a source we read
- We tested
- Not tested yet
- reported from sources; we did not run it ourselves
- Gaps we flag
- 4
- things we could not verify
- Sources
- 2
- 2 primary, linked below
What changed
| # | What the source says | Status |
|---|---|---|
| 01 | The task: given a production-level website an agent can explore through a browser, it must reverse-engineer a full-stack version matching the site's appearance and behavior, per the README. | Reported |
| 02 | Scale: 70 websites and 46,916 hidden test cases, published on Harbor Hub and Hugging Face and downloadable with the websitebench PyPI package (pip install websitebench). | Reported |
| 03 | Scoring is strict: each task has 302 to 2,016 hidden cases, and a case passes only if its functional checks pass and every visual checkpoint reaches a normalized SSIM of 0.99. A task's reward is the share of cases passed. | Reported |
| 04 | Leaderboard of eight models: Claude Opus 5.5 leads at 26.2 overall, then GPT 6.1 Sol at 20.2 and Claude Opus 5 at 17.1. The best open-weight model listed is GLM 5.3 Flash at 14.7; Qwen 3.8 Max is last at 7.3. | Reported |
| 05 | Cost spread is wide: the site reports mean resource use per task of $305.1 and 1,095.6M tokens for Claude Opus 5.5, versus $35.8 for GPT 6.1 Sol and $4.6 to $4.9 for DeepSeek V4.1 Flash and GLM 5.3 Flash. | Reported |
| 06 | Runs use the Harbor framework (the README pins harbor==0.25.0) with Docker; code, task definitions and verifier are Apache-2.0, while the reference sites are for non-commercial research only. | Reported |
reported by the sourcemeasured or run by usnot verified
Primary source: TIGER-AI-Lab/WebsiteBench on GitHub (README). We are the fast, independent index; the vendor’s own page is the source of record for this release.
Leaderboard: score versus cost per task
Figures copied from the project's leaderboard (read Nov 13, 2026). Overall is the hidden-test pass rate across the 70 tasks; cost and tokens are mean per task as the site reports them.
| Model | Overall | Mean cost per task | Mean tokens per task |
|---|---|---|---|
| Claude Opus 5.5 | 26.2 | $305.1 | 1,095.6M |
| GPT 6.1 Sol | 20.2 | $35.8 | 136.5M |
| Claude Opus 5 | 17.1 | $202.0 | 352.6M |
| GLM 5.3 Flash (open) | 14.7 | $4.9 | 154.2M |
| DeepSeek V4.1 Flash (open) | 13.5 | $4.6 | 133.1M |
| Grok 4.6 | 12.1 | $75.0 | 37.1M |
| Kimi K3 (open) | 10.6 | $36.1 | 83.7M |
| Qwen 3.8 Max (open) | 7.3 | $9.6 | 30.3M |
What the numbers suggest (our reading)
Our reading: even the leader fails about three in four hidden tests, so full website cloning by agents is far from solved. The 0.99 SSIM visual bar is strict, which likely keeps scores low even when a rebuilt site looks close to a person.
Our reading: price per point matters more here than rank. By the site's own numbers Claude Opus 5.5 spends about 8.5 times what GPT 6.1 Sol spends per task for 6 more points, and the two cheapest open models land 5.5 to 6.7 points behind GPT 6.1 Sol at about a seventh of its cost.
- The leaderboard splits scores into Request, Browser and System columns; browser-level checks are the lowest for every model listed, from 21.8 down to 4.1.
- Agent harness matters: the quickstart runs the Codex agent, and the site does not say in what we read whether every model used the same harness.
In the source’s words
“A case passes when its functional checks pass and every visual checkpoint reaches a normalized SSIM of 0.99.”
“hidden-test pass rates over the 70 tasks and mean resource use per task for eight models.”
What we have not verified
- We did not run WebsiteBench or any agent on it; all scores and costs are the authors' published figures.
- The paper is listed as a badge but we found no linked paper or preprint to check methodology details.
- We did not confirm which agent harness, settings or number of runs produced each leaderboard row.
- Licensing of the reproduced third-party sites rests on the README's non-commercial research statement; we did not review it legally.
This brief is Reported: written from the sources below, not from running the release ourselves.
Sources
- TIGER-AI-Lab/WebsiteBench on GitHub (README)primary source · fetched 2026-11-13
- WebsiteBench project page and leaderboardprimary source · fetched 2026-11-13
Related
Spotted an error? See our corrections policy or email the address on the contact page. Markdown version for agents: /news/websitebench-agents-rebuild-websites-benchmark.md
Want these briefs in your Google Top Stories? Add Agentic Weekly as a preferred source on Google.
Newsletter
The Weekly Verdict. Free, every Monday.
A free email every Monday: the week's few big AI releases, then 5 short use cases for the new models, each with a prompt you can paste.
Free. Double opt-in. One click to unsubscribe.