Skip to content
VibeFather
Sign up free

Live data

What AI coding agents build from one prompt

Every number below is counted from the 538 websites live on VibeFather. Each was built by an AI model in a real coding tool from one published prompt, then reviewed on desktop and mobile. Medians, not best cases.

Live websites
538
Templates
265
AI models
8
Multi-page builds
28%
Files per build
32
Tokens per build
~4.9M

Files and size are medians of the built output. Tokens: median of the 202 builds where the tool reported a full count, cached context included.

By AI model

ModelRan inSitesMulti-page3DFiles (median)Most built
Gemini 4 CarbonCursor153432Landing Page, E-commerce
Kimi K3.1 ProKimi Code81291323Landing Page, Interactive Toy
GPT-6.1 SolCodex2011214Landing Page, Interactive Explainer
Claude Opus 5.5Claude Code, Cursor6620633Landing Page, Interactive Explainer
Grok 4.7Grok Build2021830Browser Game, Event
Claude Sonnet 5.5Claude Code, Cursor3610635Landing Page, Interactive Explainer
DeepSeek V4.1 FlashDeepSeek Harness179571627Landing Page, E-commerce
GLM-5.3ZCode12127240Landing Page, E-commerce

Each model built a different set of templates, so these columns describe what we asked for as much as the model. For a fair comparison, open a comparison page: two models, same prompts.

What gets built

Frameworks and libraries

Share of all builds; a build can have several tags.

How the data is collected

  1. We write one prompt per template: brief, stack, variables, design system and acceptance checklist.
  2. An AI model builds the site in a real coding tool from that prompt alone, in an empty project.
  3. Every build is tested automatically (console errors, external requests, licenses) and screenshotted on desktop and mobile, then reviewed for layout, responsiveness, motion and copy.
  4. Only builds that pass go live; the model, tool, page count, built output and token count (when the tool reports it) are recorded with the build.

More detail: how we make templates.

FAQ

Building websites with AI, by the numbers

Can AI build a complete website from one prompt?

Yes. 538 live websites on VibeFather were each built by an AI coding agent from a single published prompt, with nothing added by hand, across 8 AI models and 7 coding tools. Every build was checked on desktop and mobile before it went live.

How many tokens does it take to build a website with AI?

In our builds, the median was about 4.9M tokens per website, counting the agent's cached context, across 202 builds where the coding tool reported a full count. Most of that is context the agent re-reads while it works; the output it writes is a small share. Tools count tokens differently, so compare numbers from the same tool only.

Which AI models can build websites?

So far Gemini 4 Carbon (15), Kimi K3.1 Pro (81), GPT-6.1 Sol (20), Claude Opus 5.5 (66), Grok 4.7 (20), Claude Sonnet 5.5 (36), DeepSeek V4.1 Flash (179) and GLM-5.3 (121) have live sites here. The model pages show every site each one built, and comparison pages show two models on the same prompt.

Which framework do AI coding agents use for websites?

Whatever the prompt asks for. Across our builds: React 83%, Tailwind CSS 80%, Motion 63%, Three.js 14%, Plain HTML 13% and GSAP 7%. React is the most common because many prompts ask for it.

Can AI build a multi-page website?

Yes. 149 builds are multi-page sites (28% of all builds), up to 6 pages from one prompt.

Can AI build a 3D website?

Yes. 77 builds use 3D / WebGL (14%), mostly with Three.js. They run in the browser with no server.

How big is a website an AI agent builds?

The median build ships 32 files and about 1.3 MB after the build step, images included.

Where does this data come from?

From VibeFather's own production records: every published example, the model and tool that built it, the template's tags and page count, the built output and, where the tool reports it, token usage. The numbers update as new sites go live. You may cite them with a link to this page.

Data: VibeFather, CC BY 4.0 — cite with a link to vibefather.ai/stats.