Benchmarks for how AI designs
Coding evals check if the code runs. 21st Lab checks if developers actually pick what the model made, on 21st.dev, with cost next to every score
Live Preferences · GPT-6 Astra leads at 3.4%, then GPT-6.1 Sol 3.3% and Grok 4.7 2.8%
Taste Test
Agrees with developers' picks
- 1Claude Opus 5.577%
- 2Claude Sonnet 5.572%
- 3GPT-6 Astra72%
- 4GPT-6.1 Sol71%
- 5Claude Fable 5.171%
- 6Gemini 4 Argon69%
- 7Muse Spark 1.367%
- 8Kimi K3.1 Pro66%
- 9Qwen 3.8 Max63%
- 10MiMo v2.6 Pro62%
- 11Grok 4.762%
Live Preferences
Pick rate on live traffic, hero sections
- 1GPT-6 Astra3.4%
- 2GPT-6.1 Sol3.3%
- 3Grok 4.72.8%
- 4Claude Sonnet 5.52.7%
- 5Claude Fable 5.12.6%
- 6Muse Spark 1.32.6%
- 7Gemini 3.8 Flash2.5%
- 8Claude Opus 5.52.1%
- 9Qwen 3.8 Max2.1%
- 10GLM 5.32.0%
- 11MiMo v2.6 Pro1.8%
- 12Kimi K31.1%
Animation Reconstruction
Animations matched or close
- 1GPT-6.1 Sol100%
- 2GPT-6 Astra96%
- 3Claude Opus 5.591%
- 4Claude Sonnet 5.587%
- 5Kimi K3.1 Pro87%
- 6Claude Fable 5.187%
- 7Gemini 4 Argon83%
- 8Qwen 3.8 Max52%
- 9MiMo v2.6 Pro43%
- 10Muse Spark 1.343%
- 11Kimi K335%
- 12Gemini 3.8 Flash30%
- 13Grok 4.717%
Taste Test
- Pairs from real 21st searches, where developers copied one component far more than the other
- The best, Claude Opus 5.5, agrees with developers 77% of the time. A coin gets 50%
- Where developers go for gradients, glow or 3D, the models pick the plain option
Updated 11/02/2026 · 11 models
Top three
- 1Claude Opus 5.577%
- 2Claude Sonnet 5.572%
- 3GPT-6 Astra72%
Live Preferences
- Every model designed hero sections for 21st's hero page
- Developers saw them in the feed without knowing the model, and opening one counted as a pick
- Blind duels, cost and time next to every score
Updated 11/02/2026 · 12 models
Top three
- 1GPT-6 Astra3.4%
- 2GPT-6.1 Sol3.3%
- 3Grok 4.72.8%
Animation Reconstruction
- Models rebuilt animated 21st components from screenshots alone
- Rounds of fixes, then each rebuild is judged a match, close or off against the original
- Pointer effects, shaders and other motion
Updated 10/27/2026 · 13 models
Top three
- 1GPT-6.1 Sol100%
- 2GPT-6 Astra96%
- 3Claude Opus 5.591%
Research
The studies behind every benchmark, with the method, the pairs and the clips. By the 21st team
AI models predict which UI components developers choose, up to 77% of the time
We showed models pairs of components from 21st and asked which one developers chose more often. Claude Opus 5.5 did best, and all of them prefer the plain option where developers go for gradients, glow and 3D.
We tested 12 AI models on hero section design with real users
We mixed AI-made hero sections into the 21st feed and measured which ones developers open.
Can AI models rebuild an animation from screenshots?
We gave 13 AI models screenshots of animated components from 21st and asked for the code. GPT-6.1 Sol and GPT-6 Astra rebuilt nearly all of them, and shaders split the rest.
How Claude Code uses a component catalog differently from humans
Three months of searches on 21st from Claude Code and from humans on the site: what each one searches for and what it picks.
Have a design task you want models tested on?
We run it on 21st.dev, where developers pick UI for their own projects, and show cost next to every score
Talk to us