Methodology
How 21st Lab measures what AI designs, on the library developers use every day.
Built on 4.4M+ users and 1.16M+ registered accounts on 21st.dev.
- Users
- 4.4M+
- Unique browsers since Dec 2024, GA4 ¹
- Registered accounts
- 1.16M+
- As of Oct 30, 2026
- Component views
- 65M+
- Last 30 days ²
- Took code
- 115K+
- Developers who copied, installed or generated code in the last 30 days ³
Benchmark format
How a benchmark runs
- 01
One brief
A design task written like a real request: 4 hero briefs, from a SaaS product hero to a cinematic dark one
- 02
12 models, one prompt
Every model gets the same system prompt, the same allowed libraries and no external assets, at its highest reasoning setting that finishes the job
- 03
Rendered and checked
Each output is rendered at 375 and 1440 px, light and dark. A model gets another attempt only when its first one doesn't work
- 04
Published blind
Sections go out from one neutral account with matching covers. The card never shows the model
- 05
Live in the feed
Signed-in developers see two sections from the same brief, by two different models, at fixed slots in the real feed. 10% get the normal feed as a control
- 06
Picks become a ranking
Opening a section is a pick. Pick rates get a 95% interval, and every view where exactly one of the two was opened is a duel for a Bradley–Terry rating
Every page view where a developer sees both sections is a blind comparison between two models on the same task. Briefs, pairs and slots are drawn at random, so every model appears equally often in each slot and position evens out across the ranking.
Why this matters
Honest evaluation, where the work happens
Real choices, not votes
Nobody is asked to judge. Developers pick UI for their own projects, and the pick is the signal
Taste, measured
Design quality is whether people want it. We count what they open and copy, on the page where they already shop for UI
Same slot, same cover
Position and presentation are fixed and randomised, so the only thing that changes between two cards is the model
Cost next to every score
Each result carries what it cost and how long it took, so a cheap model that is just as liked shows up as one
Ranking calculation
From picks to a ranking
Pick rate. The share of members who opened a model's section after seeing it. Each member counts once per model, and the rate is adjusted for the mix of briefs each model was shown. The interval is a 95% Wilson interval.
Head to head. Only page views where a member saw both sections and opened exactly one count; the opened one wins. A Bradley–Terry model, fitted by Newton's method with a light ridge, turns those wins into ratings on the Elo scale (1500 is the average model) and takes out the advantage of the higher slot.
Ties. A model whose interval reaches the leader's is reported as tied with it. We don't call a winner the data can't separate.
Technical standards
What stays fixed
Models run at their highest reasoning setting that still finishes the job, with one shared system prompt: React 19 and Tailwind v4, a short list of allowed libraries, no external images, fonts or URLs, and a layout that works from 375 px wide.
Cards are published from one neutral account with matching covers and hidden counters, so an early lucky winner can't compound its lead. Admins and the publishing account are excluded from every count. The unit is the member, not the event.
Every number on a benchmark page links back to how it was measured. Questions about the method: talk to us.
Our users
Who picks the designs
21st.dev is where developers and designers shop for UI: they browse, open and copy components into their own projects, mostly as prompts for a coding agent. Every benchmark runs on that traffic, so the people picking designs are the people who build with them.
- Users
- 4.4M+
- Unique browsers since Dec 2024, GA4 ¹
- Registered accounts
- 1.16M+
- As of Oct 30, 2026
- Component views
- 65M+
- Last 30 days ²
- Took code
- 115K+
- Developers who copied, installed or generated code in the last 30 days ³
- Paying members
- 12,600+
- Active subscribers, Oct 30, 2026
- Typed searches
- 1.5M+
- Last 30 days
- Page views
- 75M+
- Since Dec 2024, GA4 ¹
- Components in the library
- 7,000+
- Public components, Sept 2026 sitemap
- Countries and territories
- 200+
- Last 12 months, GA4
- Engineer28.5%
- Designer27.8%
- Student17.3%
- Founder9.4%
- Product4.9%
- Marketer4.3%
- Other7.8%
“What best describes you?”, answered by 421,722 of 1,160,780 registered users ⁴
- Personal projects50.8%
- Work36.9%
- Education12.3%
Use case, answered by 414,558 registered users
- YouTube25.9%
- Instagram24.3%
- Google search12.8%
- TikTok9.1%
- Twitter / X8.1%
- Friend or colleague5.5%
- ChatGPT and AI assistants4.7%
- Facebook3.1%
- Reddit1.9%
- Other4.6%
Answered by 398,397 registered users
- Landing page sections17.7%
- Full landing pages and templates13.1%
- Cards, media and galleries12.8%
- Backgrounds, shaders and effects11.8%
- Hero sections10.7%
- Navigation and layout9.1%
- Dashboards and data viz7.4%
- Buttons, inputs and forms6.4%
- Feedback and overlays3.1%
- AI chat and agent UI3.1%
- Other4.8%
Page views of /s/<tag> pages (2.06M, 22.5% of all page views), grouped by topic
- Browse feed26.0%
- Search24.8%
- Tag pages24.5%
- Home showcase8.3%
- Landing page6.7%
- Similar components6.4%
- Author profiles2.0%
- Other1.3%
65.3M component views by the surface that led to them
- Asia43.6%
- Americas26.8%
- Europe22.6%
- Africa5.0%
- Oceania1.5%
- Not set0.6%
- India13.9%
- United States13.1%
- Brazil7.2%
- Germany3.7%
- France3.0%
- China2.9%
- United Kingdom2.5%
- Türkiye2.3%
- Indonesia2.0%
- Rest of the world49.4%
Singapore (6.8%) is folded into the rest: its traffic looks automated (about one session per user)
- Windows44.9%
- macOS30.6%
- iOS12.3%
- Android8.3%
- Linux3.3%
- Other0.7%
Operating system. Devices: desktop 79%, mobile 21%
- ChatGPT81.8%
- Gemini11.4%
- Claude4.3%
- Perplexity2.2%
- Copilot and Grok0.2%
About 50K users who arrived from an AI assistant with a referrer; many more arrive without one and count as direct
- 1.GA4 since 2024-12-02 (the site launched about two months earlier; that traffic is under 0.3% of the total). GA counts browsers and cookies, not people, and ad blockers hide part of a developer audience.
- 2.Every component view event in the 21st database over the last 30 days, split by the surface that led to it.
- 3.Unique developers who left with code over the last 30 days: a copy on the site, get_component or generate over MCP, get, add or generate over the CLI.
- 4.Answers merged from every version of the sign-up questionnaire into one set of roles; 739K registered users skipped the question.