Google Stitch Rapid: Interactive Interfaces in 5 Seconds Powered by Gemini Decisions APINote · digest for November 15
Set of interactive mobile screens with buttons and toggles generated in Google Stitch Rapid
Note4 sections≈ 4 min

Google Stitch Rapid: Interactive Interfaces in 5 Seconds Powered by Gemini Decisions API

When you need to show a team or client how a new mobile app will work—say, a restaurant table-booking flow—you usually face two awkward options. You can drag rectangles around a design tool to produce static mockups with no real transitions, or ask a chatbot to code the layout and watch a blinking cursor type out markup line by line.

On November 12, Google Labs offered a third path by launching the experimental Stitch Rapid mode on top of the Gemini Decisions API. It turns a short spoken or typed prompt into a clickable prototype mid-conversation. What trick lets it assemble and restyle screens almost instantly?

The Prototyping Speed Barrier: Why Standard LLMs Lag on UI Layout

Traditional language models are slow at UI because of how they output text. To render a screen, a standard model acts like a stenographer, sequentially printing HTML, React, or Flutter code token by token—every tag, class, and style included.

Imagine building that restaurant app: date and time pickers, an interactive floor plan, and an SMS confirmation screen. Writing that markup character by character takes a standard model 40–90 seconds. And if the restaurant owner asks on the call to round the booking button or drop the floor plan, the generator has to rewrite the whole block from scratch.

Gemini Decisions API Architecture: Picking Ready-Made Components Instead of Generating Text

To bypass that bottleneck, Google engineers built the rapid mode on gemini-decision-1, which Julian Goldie broke down in his Stitch update review. I checked that breakdown against Google's spec, and one detail jumped right out: the model runs in zero output tokens mode—instead of writing text letter by letter, it operates as a classifier.

Three mechanics drive the speed:

  • Parallel routing: ingesting up to 256K tokens of context, the model answers 1,000 parallel questions in a single pass. Like an art director flipping through a catalog, it selects a navigation pattern for the restaurant, a calendar widget, and one of 7 built-in design systems (Material 3, iOS HIG, Minimalist, SaaS Dashboard, etc.), returning probabilities and confidence scores.
  • Quarter-second decisions: picking pre-built UI components takes 150–250 milliseconds, which is why a multi-screen clickable wireframe (Playable Wireframe) compiles from a sketch or short phrase in 5 seconds.
  • Pennies-on-the-dollar pricing: in Google AI Studio, calls cost $0.10 per 1 million input tokens and $0 for output decisions (the same router also dispatches requests across models in Gemini Notebook).

Working Directly on the Canvas: Taps, Sliders, and Voice Edits Without a Prompt Box

Switching to classification reshapes the interface itself. When the engine makes decisions in milliseconds, putting a chat box between you and the mockup just gets in the way—so every control moves onto the canvas.

In the restaurant prototype preview, clicking an unlinked button triggers Design Next Screen to build the following view on the fly. Don't like a table card? Tap it for fresh variations in 2 seconds, or right-click to open the Tuner panel with spacing and typography sliders. Need to rethink the booking step? Hold the spacebar like a walkie-talkie and say, "Swap the card grid for a horizontal carousel and make the order button round"—and the screen updates immediately.

Moving to the Main Stitch Editor and Honest MVP Limits

You can open the canvas at stitch.withgoogle.com/rapid (I covered the core platform in an earlier deep dive on streaming UI generation in Google Stitch). There is one major catch with this MVP: drafts don't persist between sessions. Close the tab, and your restaurant prototype is gone. As soon as you've tested an idea, download it or hit Send to Stitch to move it into the main Google Stitch editor for polish, HTML/CSS export, and Figma handoff (direct .fig sync via the Figma API is still in the works).

There's still no single tool that handles both instant napkin sketches and a production design system: Stitch Rapid won't build custom animations, and it doesn't forgive closed tabs. Still, now you know the rules—and you can turn a rough pitch into a live, clickable prototype right in the middle of a client call.

Want a deep diveVotes help us choose what to break down in detail. This material has 0 votes.
Sources · 2