Google Stitch Runs on Gemini 3 Now, and It’s Still Free for a Little Longer

Google Stitch launched in March 2026 running on Gemini 2.5. It does not run on Gemini 2.5 anymore. Google has since upgraded the tool to Gemini 3, added a way to chain screens into a working prototype, and kept the whole thing free inside Google Labs. That last part has a shelf life. Analysts and Google’s own Labs pattern point to Stitch leaving the free preview and picking up paid tiers around the end of 2026, which is now weeks away. If you have been meaning to test it, the version worth testing is the current one, and the price is the one detail most likely to change first.

This is a refreshed look at where Stitch actually stands, written for creators who want to design an app screen, a landing page, or a client mockup without opening a design tool they do not know how to use.

Vibe design, in one paragraph

The core idea has not changed since launch. Instead of dragging rectangles around a canvas, you describe what a screen should do and how it should feel, and Stitch generates the interface. You can type the prompt, upload a screenshot or sketch as a reference, draw notes directly on the result, or talk to it through the voice canvas. It grew out of Galileo AI, a startup Google acquired and folded into Labs, and it inherited that tool’s text-to-UI approach before Google rebuilt it on Gemini. The pitch is the same one that made it interesting in the first place: you do not need to think in components to get a usable screen out of it.

What Gemini 3 actually changed

The model swap is the real update, not a cosmetic version bump. By mid-2026 Stitch was running on Gemini 3, and the tool now splits into two modes. Standard mode uses a fast Gemini 3 Flash model for quick iteration, and a slower thinking mode uses a Gemini 3 Pro model when you want higher-fidelity output and more careful layout reasoning.

In practice the jump from 2.5 to 3 shows up in multi-screen coherence. Ask for an onboarding flow plus a dashboard plus a settings page, and the screens come back sharing a visual language instead of looking like three unrelated mockups. Prompt adherence tightened too, which matters most when you are describing a specific structure (“a two-column pricing page with a comparison table”) rather than a vibe. It is still an AI generator, so it has house defaults it drifts toward, but the floor is higher than it was in March.

The three features the original guide never mentioned

If you read about Stitch when it launched, three capabilities have landed since that change how it fits into a real workflow.

Prototypes. You can now stitch individual screens together on the canvas into a clickable prototype, wiring buttons to destinations and mapping an entire user flow. Google still labels this experimental, but it moves the tool past static mockups toward something you can hand a tester and say “click through this.”

Annotate. Instead of only re-prompting in text, you can draw on a generated screen directly. Circle an element, scribble a note like “make this bigger” or “this feels cluttered,” and Gemini reads the markup and applies context-aware changes. For anyone who thinks visually, pointing at the problem beats describing it.

Theme. A sidebar now manages global design tokens: light and dark mode, a primary color palette, corner radius, and font choices that cascade across every screen. The catch is that this is an auto-generated design system, not your brand kit. Stitch will keep a set of screens internally consistent, but it does not ingest your existing colors and type unless you rebuild them by hand.

Where your designs go when you are done

Export is where the tool got noticeably more useful, and where creators should read the fine print. Stitch now exports to HTML/CSS, Tailwind, Vue, Angular, Flutter, and SwiftUI, which is a real jump from the narrower web-only output it started with. It also pushes into Figma and into Google’s own AI Studio, and it connects to coding agents through an MCP server so you can send a design’s context straight to a tool like Claude Code and have it build the front end to match.

The honest caveat is direction. Some of the smoothest handoffs point back into Google’s own ecosystem, and export fidelity is best when a design was generated from a text prompt rather than from an uploaded image. Treat the code as a fast starting scaffold, not a finished front end, and plan to clean it up.

How it compares to the tools creators already use

Stitch is not trying to be a full app builder, and that is the most important thing to understand before you pick it. It designs and prototypes the front end. It does not stand up a database, handle auth, or deploy.

The clean read: reach for Stitch when you are exploring what a screen should look like and you want a prototype fast. Reach for v0 when you already know the component you need. Reach for the vibe-coding tools when you need a working back end behind the interface. Stitch handles the first stretch of the road, roughly concept to clickable prototype, and then you hand off. That handoff is where a vibe-coding workflow picks it up.

The free window, and why I would export early

Stitch is free right now, with generation caps of roughly 350 standard generations and 200 thinking-mode generations a month, where a multi-screen flow counts as a single generation. Google has not announced pricing, but the widely reported expectation is that Stitch exits Labs around Q4 2026, that the free tier shrinks to something closer to 50 to 100 generations a month, and that paid plans land somewhere around $10 to $15 a month, deliberately under Figma’s per-editor price.

Here is the operator note. Twenty-plus years running IT operations taught me to treat any free vendor preview as a dependency I do not control, and to plan for the day the terms change, because they always do. A Labs experiment is exactly that: generous today, repriced or retired on Google’s schedule, not yours. So use the current limits while they are wide, but export what you make as you go. Pull your screens down as code and Figma files, keep them in your own storage, and do not let a client project’s only copy live inside a preview product. The design work is yours; make sure the files are too.

Where Stitch earns its place, and where it does not

After the Gemini 3 upgrade this is a genuinely good ideation tool, and it is still not a substitute for design judgment. The known rough edges are worth naming before you build a workflow around it.

Generated screens often miss basic accessibility, including weak color contrast and small touch targets, so anything public-facing needs an audit before it ships. Output tends to converge on a handful of layout patterns once you have generated a few dozen screens, so distinctive work takes deliberate prompting. The layouts are static at set breakpoints rather than truly fluid, so responsive behavior is manual work after export. It is single-user, with no live collaboration. And as LogRocket’s review put it bluntly, leaning on Stitch alone “will almost certainly result in a product riddled with usability issues.” Read that as a boundary, not a dismissal.

The workflow that works: use Stitch to get from a vague idea to a clickable prototype in an afternoon, test it with five or so real users (the Nielsen Norman Group found that many uncovers around 85% of usability problems), then export and hand the result to a coding tool or a developer to build for real. For a creator who has an app idea and no design budget, that is a legitimately useful first mile, and right now it costs nothing. Both of those things are worth acting on before the pricing page arrives.