/
Copy Bench

Copy Benchv1

Copy Bench measures how well models rebuild complex web components from a screen recording: WebGL, shaders and multi-step interactive animations.

Read the full write-up

Nov 15, 2026 · by 21st Labs

Leaderboard

1GPT-6 Astra0.0
2GPT-6.1 Sol0.0
3Claude Fable 5.20.0
4Gemini 4 Carbon0.0
5GPT-6 Sol0.0
6Claude Opus 5.50.0
7Qwen 4 Max0.0
8Muse Wonder 1.00.0
9Qwen 3.8 Max Prime0.0
10Kimi K3.1 Proopen weights0.0
11MiniMax M3open weights0.0
12Grok 4.70.0
13Claude Sonnet 5.50.0
14MiMo V2.6 Proopen weights0.0
Time per copy
Minutes per copy, fastest first. Solid: one-shot, faint: added by self-verify
54s
2.3
2.6
6.9
8.1
8.7
19.0
20.4
21.0
21.9
23.7
39.6
42.5
43.1
Cost per copy
Dollars per copy, cheapest first. Solid: one-shot, faint: added by self-verify
$0.10
$0.29
$0.44
$0.56
$0.84
$0.90
$0.90
$1.27
$1.41
$1.74
$1.84
$3.62
$4.74
$8.70

Side by side

Neon Loop
Original: Neon Loop
Original
Claude Opus 5.5
Warp Carousel
Original: Warp Carousel
Original
Claude Opus 5.5
LED Badges
Original: LED Badges
Original
Claude Opus 5.5
Distressed Denim
Original: Distressed Denim
Original
Claude Opus 5.5
Volumetric Light Shafts
Original: Volumetric Light Shafts
Original
Claude Opus 5.5
Softbox Character
Original: Softbox Character
Original
Claude Opus 5.5
Fur Ball
Original: Fur Ball
Original
Claude Opus 5.5
Liquid Metal
Original: Liquid Metal
Original
Claude Opus 5.5
SVG Morph
Original: SVG Morph
Original
Claude Opus 5.5
Studio Landing
Original: Studio Landing
Original
Claude Opus 5.5
Foundation Primitives
Original: Foundation Primitives
Original
Claude Opus 5.5
Neon Hello
Original: Neon Hello
Original
Claude Opus 5.5
Silk Spider Cursor
Original: Silk Spider Cursor
Original
Claude Opus 5.5

Insights

The browser helped most copies

Self-verifying raised the score on 449 of 699 copies, left 139 within 2 points and lowered 111.

Neon Loop·GPT-6 Astra
Original: Neon Loop
Original
One-shot
Self-verified
Neon Loop, GPT-6 Astra: the browser fixed the split pink and blue background.
Warp Carousel·Claude Opus 5.5
Original: Warp Carousel
Original
One-shot
Self-verified
Warp Carousel, Claude Opus 5.5: the browser changed pretty much nothing.

Models can copy a web page. Motion and 3D are where they break.

One-shot, the average model scores 59 on pages and scroll, 41 on shaders and 35 on 3D, and self-verifying adds 7 to 10 points in every category. Self-verified, six models clear 75 on pages, but only GPT-6 Astra stays above 80 in all five.

0100
ModelPagesGames2D animationShaders3D
GPT-6 Astra90.084.572.475.170.4
GPT-6.1 Sol88.182.776.272.465.9
Claude Fable 5.279.973.872.949.348.1
Gemini 4 Carbon67.747.762.356.650.3
GPT-6 Sol77.464.864.046.545.3
Claude Opus 5.566.857.452.139.745.5
Qwen 4 Max60.845.556.943.025.7
Muse Wonder 1.060.752.642.135.027.9
Qwen 3.8 Max Prime57.049.444.738.424.2
Kimi K3.1 Pro47.342.043.028.718.8
MiniMax M337.049.741.226.619.1
Grok 4.739.740.937.122.918.6
Claude Sonnet 5.539.332.731.523.919.1
MiMo V2.6 Pro16.917.419.113.515.9
Average of 1459.252.951.140.835.4
Score per model and category, for the step picked above. Categories have 7 to 14 components, so gaps under ~5 points are noise.

On the judges' checklist, motion averages 33 out of 100, while typography gets 76.

One-shotAdded by self-verify
Typography0.0
Content0.0
Layout0.0
Color0.0
Motion0.0
Each checklist item over all 700 self-verified copies, both judges averaged. Copied or empty copies are left out.

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.

Before the change · never shown
Before the change · never shown
What the model saw
What the model saw
Claude Opus 5.5's copy
Claude Opus 5.5's copy
Opus recognized the three.js example, called the gold a rendering glitch and loaded our lamp file as is.

How we measured this →