/
Turing UI Bench

Benchmarks / Research

Turing UI Bench: Do real developers pick AI-made UI?

November 4, 2026

How often developers opened AI-made and designer-made UI
% of viewers who opened it. Designer columns: the top, median and lowest designer-made component in the same feed.
TL;DR

How does AI-made UI compare to work by real designers? Now we have an answer.

We had 12 frontier models generate 48 UI components and mixed them into the 21st feed, where more than 4 million developers browse and grab UI made by real designers. Then we tracked how people actually engaged with them.

Result: the best models are getting close, but the best designers are still ahead.

  • Top models beat the median designer. GPT-6 Astra (3.4%) and GPT-6.1 Sol (3.3%) got opened more often than the median designer-made component (2.7%).
  • The best designers are still ahead. Their top components got opened about 40% more often than the best model's.
  • The average AI component is close behind. At 2.3%, it sat just below that median.

Why another design benchmark

Most AI benchmarks check things you can verify: the code compiles, the tests pass, the math adds up. Design isn't like that. No test can tell you a landing page is good. The only real judge is people.

That's where design arenas come in, and they did something great: they put human taste at the center of how we judge AI-made UI. Two designs, one vote, and a clear ranking of models.

We wanted to ask the next question. Not "which model looks better?" but "would a developer actually ship it?"

Design arenasTuring UI Bench
The questionWhich model's design do people prefer?Do developers pick AI-made UI for their own product?
Who judgesPeople comparing models side by sideDevelopers who don't know it's AI
Next toOther modelsWork by real designers
Best forRanking modelsSeeing how AI-made UI does in the wild

Arenas tell you who's winning. This tells you if it's ready for real work.

The models and the prompt

We used 12 models, each at its highest reasoning setting that still finished the job: GPT-6.1 Sol, GPT-6 Astra, Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5, Gemini 3.8 Flash, Grok 4.7, GLM 5.3, Qwen 3.8 Max, Kimi K3, MiMo v2.6 Pro and Muse Spark 1.3.

Every model got the same system prompt. It asks for two files in the format 21st publishes (a component and a demo), React 19 with Tailwind v4, a short list of allowed libraries (Framer Motion, GSAP, Three.js, Lucide, a few shadcn/ui primitives), no external images, fonts or other URLs, so every visual is drawn in code, and a layout that works from 375 px wide with no console errors.

Each model got one plain API request per brief, with the system prompt and the brief in it. We didn't wrap the models in an agent harness, so a model never saw how its code rendered and couldn't go back to correct it.

System promptThe same for all 12 modelsShow
You are building one hero component for the 21st.dev component marketplace.

Output exactly two fenced code blocks and nothing else:
1. `component.tsx` — the reusable component. Start with "use client". Type the props with a Props type. Export the component both as a named export and as default. Accept a `className` prop and merge it with `cn()`. Every prop has a default, so `<Component />` renders the full hero with no props.
2. `demo.tsx` — imports the component from "./component" and renders it full-bleed: wrap it in `min-h-screen w-full` and nothing else. No captions, headings, control panels or "made with" badges in the demo.

Stack: React 19 + TypeScript, Tailwind CSS v4. You may import: react, framer-motion (or motion), gsap, lucide-react, three / @react-three/fiber / @react-three/drei, and standard shadcn/ui primitives from "@/components/ui/<name>" (button, badge, card, input, separator, tabs, sheet, navigation-menu) and `cn` from "@/lib/utils". Anything else must be implemented inside component.tsx.

No external URLs of any kind: no remote images, fonts, videos, 3D models, Spline scenes or CDNs. The preview runs in a sandbox that blocks every cross-origin request, so anything external renders as broken. Draw visuals in code (CSS, SVG, canvas, WebGL), and use plain text for logos.

Theme:
- If the brief says "theme: follows site theme", use only semantic tokens (bg-background, text-foreground, bg-card, text-muted-foreground, border-border, bg-primary, text-primary-foreground, …) or `var(--color-*)`; it must look right in both light and dark. Never write `hsl(var(--x))` or `oklch(var(--x))`.
- If the brief says "theme: dark", the hero paints its own dark background and must look the same whatever the site theme is.

Must work from 375px to 1920px wide with no horizontal scroll. Animations must respect `prefers-reduced-motion`. Give interactive elements stable ids. No console errors.

How the test works

The 48 sections are published from one neutral studio account with matching covers, and the cards don't show the model. Signed-in visitors to the hero page see two of them in the feed, at positions 6 and 12. Both come from the same brief and from two different models, so every page view is a blind comparison between two models. A random 10% of visitors get the normal feed as a control.

The top of the hero feed during the test
% of viewers who opened each card. Human-made cards are corrected for position to slots 6 and 12, so every card is on the same scale.

The four briefs and what the models made

Ledger

A SaaS product hero (dark):

Build a hero for "Ledger", invoicing software for freelancers. Headline "Get paid without the chasing", subheadline "Send invoices in seconds, track every payment, and let reminders do the awkward part.", call to action "Start free". Everything else (layout, visuals, colors, typography, motion) is up to you.

Aurelia

A scroll-expanding media hero (dark):

Build a hero for "Aurelia", a travel film studio. A framed media panel sits in the center of the first screen with the title "Stories from the edge of the map" split around it. As the user scrolls, the panel grows until it fills the viewport, the two halves of the title slide apart and fade, and the page then continues to a short paragraph: "Twelve films. Six continents. One crew that never stops moving." The media panel is a slow, abstract animated landscape drawn in code. Mood: cinematic, quiet.

Orbit

An interactive 3D hero (follows the site theme):

Build a hero for "Orbit", a satellite-data API. Left: eyebrow "Now in public beta", headline "The planet, as an API", subheadline "Query imagery, weather and ground truth for any coordinate in a single call.", buttons "Get an API key" and "Read the docs". Right: an interactive 3D globe or object that slowly rotates and tilts toward the cursor. On mobile the object moves above the text. Mood: precise, technical.

Nocturne

An atmospheric dark hero:

Build a hero for "Nocturne", a sleep-sound app. A dark, atmospheric scene with a slow light effect (beams, glow, aurora or stars, your choice) behind very large type: "Fall asleep to something beautiful". Below: "Generative soundscapes that never repeat, tuned to how you breathe." and buttons "Try it free" and "Listen to a sample". Elements reveal in sequence on load. Mood: dreamy, slow, premium.

How the sections did

Cards higher in the feed get opened more, so we compare each AI section with the human card that the control group saw in the same slot. Both sides count only the default view of the hero page, and each member once.

How often a card is opened, by its position in the feed
% of viewers who opened the card.
Human-made cards (control)AI-made hero sectionsControl average 3.8%
PositionAI-madeHuman-made
63.8%3.1%
122.7%2.8%

In slot 6 the AI-made heroes did a little better, in slot 12 a little worse. Neither gap is bigger than the noise.

They didn't put anyone off the page either. 55% of developers who got the AI-made heroes opened something on the page, against 52% in the control. 20% copied something, against 19%.

And the novelty didn't wear off. The gap between AI-made and human-made heroes stayed about the same in each of the three weeks.

The brief mattered too, though less than the slot.

How often AI-made heroes are opened, by brief
% of viewers who opened an AI-made hero, all 12 models together.

Aurelia and Orbit got opened noticeably more than Ledger and Nocturne. Aurelia is also the brief where the models' designs differed the most. On the other three, most models landed on the same layout.

What happens after the click

Opening a preview is curiosity. What comes next is closer to a decision: copying the component, trying to copy it and hitting the free limit, bookmarking it, or going to its page. We counted each of these when the same developer did it within 30 minutes of opening a hero.

What developers did after opening a hero
% of developers who opened the hero. Copied it includes trying to copy and hitting the free limit.
AI-madeHuman-made, slots 4–12

Taken together, AI-made and human-made heroes look the same after the click. 32% of the developers who opened an AI-made hero did at least one of these things, against 30% for the human-made heroes around them.

The mix differs a little:

  • Copied about as often. Counting everyone who copied or hit the free limit trying to, it's 8.0% of opens for AI-made heroes, against 9.5%. Copies alone were a bit lower, 6.3% against 8.3%.
  • Bookmarked about twice as often. 4.6% of opens, against 2.4%.
  • Closed right away less often. 14% moved on to another card within 5 seconds, against 18% for human-made heroes. Both kinds of preview load in the same time, so it isn't waiting.

The free-limit paywall doesn't say which component it was for, so we count a hit when it came after a developer opened a hero and before they opened the next one.

Which models won?

The models do differ, but no single leader is ahead of the rest beyond the margin of error.

Pick rate is the share of developers who opened a model's hero after seeing it. Take Claude Fable 5.1: 3,005 developers saw at least one of its four heroes, and 79 of them opened the one they saw, so its pick rate is 2.6%. Some briefs get opened more than others, so the rates below are adjusted for each model's mix of briefs.

How often developers opened each model's hero
% of viewers who opened the model's hero.

GPT-6 Astra and GPT-6.1 Sol are on top, but they aren't clear of the next few models. Kimi K3 is clearly last. Price doesn't buy a better result: GPT-6.1 Sol cost $1.50 for its four heroes and Claude Fable 5.1 cost $15.96, and Gemini 3.8 Flash, at $0.22, lands in the middle of the pack.

The spread between heroes is wide. Grok 4.7's Aurelia, the most opened hero overall, was opened by 4.9% of the developers who saw it. MiMo v2.6 Pro's Ledger, the least opened, was opened by 0.5%, and Qwen 3.8 Max's Ledger by 0.8%. Even within one brief the gap is large: GLM 5.3's Aurelia was opened by 1.4%.

Want a study run on your model? Talk to us