21st Lab · Animation Reconstruction
13 models rebuilt animations from screenshots
Shaders, mouse effects and canvas pieces from the 21st catalog. Each model saw frames only, no code and no name, then got two rounds to compare its render with the original
Designed and run by the 21st team
GPT-6.1 Sol rebuilt every one, and was the cheapest of the six models that got 87% or more
13 models · sorted by final verdicts
| Model | Matched or close, first → final | Cost |
|---|---|---|
GPT-6.1 Sol | 83% → 100% | $24.59 |
GPT-6 Astra | 96% → 96% | $138.48 |
Claude Opus 5.5 | 70% → 91% | $59.02 |
Claude Sonnet 5.5 | 61% → 87% | $25.89 |
Claude Fable 5.1 | 74% → 87% | $215.22 |
Kimi K3.1 Pro | 70% → 87% | $27.10 |
Gemini 4 Argon | 65% → 83% | $37.52 |
Qwen 3.8 Max | 39% → 52% | $17.73 |
Muse Spark 1.3 | 39% → 43% | $9.92 |
MiMo v2.6 Pro | 35% → 43% | $1.79 |
Kimi K3 | 35% → 35% | $16.13 |
Gemini 3.8 Flash | 26% → 30% | $12.27 |
Grok 4.7 | 22% → 17% | $42.45 |
| Blank page (control) | 0% | – |
GPT-6.1 Sol
$24.59
100%GPT-6 Astra
$138.48
96%Claude Opus 5.5
$59.02
91%Claude Sonnet 5.5
$25.89
87%Claude Fable 5.1
$215.22
87%Kimi K3.1 Pro
$27.10
87%Gemini 4 Argon
$37.52
83%Qwen 3.8 Max
$17.73
52%Muse Spark 1.3
$9.92
43%MiMo v2.6 Pro
$1.79
43%Kimi K3
$16.13
35%Gemini 3.8 Flash
$12.27
30%Grok 4.7
$42.45
17%- Blank page (control)0%
What the scoreboard says
One model got every component
100%
GPT-6.1 Sol, matched or close after the fixes
It “got every component right or close after the fixes, and it was the cheapest of the six models” that got 87% or more
Price doesn't buy a better result here
$215.22
Claude Fable 5.1, for 87%
GPT-6.1 Sol got 100% for $24.59, and Claude Sonnet 5.5 matched Fable's 87% for $25.89
Fixes help almost everyone
22% → 17%
Grok 4.7, the only model that lost ground after the fixes
Most models gained after seeing their own render. A blank page, as a control, scores 0%
Then watch it
Each model got screenshots only, no code and no name, then two rounds where it saw its own render next to the original
Radial Shader: the original next to four models'every model's final version. “GPT-6.1 Sol and GPT-6 Astra match it, blade for blade” All 13
Graded by eye from one frame per component: match means it looks the same, not that it moves the same
Animation Reconstruction
Seeing its own render
Claude Opus 5.5 first drew rings, then replaced them with blades after seeing its render next to the original
16 → 21
Claude Opus 5.5, matched or close, first answers vs after the fixes
61% → 87%
Claude Sonnet 5.5, which “caught up with Claude Fable 5.1, which cost eight times more”
67% → 100%
GPT-6.1 Sol on shaders, where seeing their own render helped most
GPT-6 Astra had 96% both times: its first answers were already right, and the fixes repaired one component and broke another
Blank page (control): 0%
The same data, licensed
Components + screenshots
Real catalog components and their frames
→ used in Motion bench
Evals
Briefs, references and graded renders
→ used in both benches
The 21st Lab offer: we can run your model through the motion bench's setup and keep the results private
Put your model through the motion bench
Tell us what your model should learn. We send a sample in the format of the full delivery
Talk to usHow we ran it+
- 12 models, “each at its highest reasoning setting that still finished the job”
- Every output rendered in a headless browser at 375 and 1440 px, in light and dark mode
- “The whole run came to about $50 including failed attempts”
- “There were 657 duels, at least 86 for every model”
- Head to head is a Bradley–Terry model “on the Elo scale, where 1500 is the average model”
- 13 models on mouse effects, shaders and other motion, with render-and-revise rounds
- “Background Beams With Collision is the only component all 13 models got. It is a widely used open-source component”, a hint of memory over rebuilding
- “None of the more than 1,100 files the models wrote embeds or downloads an image”
- Verdicts were judged by eye from one frame per component (t = 4.6 s): “a single frame cannot show speed, timing or mouse interaction, so ‘match’ means visually, not behaviourally”
Designed and run by the 21st team. Write-up coming soon