Benchmarks






























Copy Benchv1
Copy Bench measures how well models rebuild complex web components from a screen recording: WebGL, shaders and multi-step interactive animations.
Nov 15, 2026 · by 21st Labs
Leaderboard
Side by side













Insights
The browser helped most copies
Self-verifying raised the score on 507 of 799 copies, left 168 within 2 points and lowered 124.


Models can copy a web page. Motion and 3D are where they break.
One-shot, the average model scores 64 on pages and scroll, 46 on shaders and 41 on 3D, and self-verifying adds 7 to 9 points in every category. Self-verified, eight models clear 75 on pages, but only GPT-6.2 Astra and GPT-6.2 Sol and GPT-6 Astra stay above 80 in all five.
On the judges' checklist, motion averages 39 out of 100, while typography gets 79.
The leaders check their work without a browser
With no browser, GPT-6 Astra wrote its own checks on every component and revised 9.9 times; Grok 4.7 revised 1.2.
More time doesn't buy a better copy
Claude Fable 5.2 scored 69.0 in 2.6 minutes; Gemini 4 Carbon took 42.5 for 67.2.
Copied or remembered?
Shown 10 tweaked famous components, Claude Opus 5.5 copied 90% of the visible changes, but wrote the original's value for 19 of 28 hidden ones.
Take the lamp from the three.js iridescence example: we made it gold and sped up its orbit 4x, and from a screenshot Opus still wrote the original's exact spin speed. More on this in the next post.


