← Blog

Google's Decisions API Lets You Ask AI a Thousand Questions at Once

Google just shipped a new model and API endpoint built for one thing: fast, confident yes/no and multiple-choice judgments across text, images, audio, video, and PDFs. Here is what that actually means for designers who are starting to build with AI.

By ProbG · November 14, 2026

Google just shipped the Gemini Decisions API, powered by a new model called Gemini Decision 1, and it does something the standard chat-style AI endpoints do not: it answers up to 1,000 structured questions about a single piece of content in one shot, returning a confidence score for every answer. If you are building anything with AI that needs to make quick, reliable calls based on what it sees or hears, this is worth understanding.

What is actually different here

Most AI APIs work by generating text token by token. You ask a question, the model starts typing, and you wait. That is fine for drafting copy or explaining a concept, but it is slow and expensive when all you really need is a judgment: is this image appropriate? Which category does this belong to? How severe is this problem?

The Decisions API skips the text generation step entirely. Gemini Decision 1 scores your candidate options in a single pass and returns probabilities and confidence scores in roughly 150 to 250 milliseconds end to end. According to Google, that makes it nearly five times cheaper than their standard Gemini Flash model for the same classification tasks, at $0.10 per million input tokens with no output tokens billed at all.

The three question types are worth knowing by name:

You can mix all three in a single request, against a single shared input, whether that input is a photo, a PDF, a video clip, or a combination of all of them.

Why a designer building with AI should care

Here is the thesis: this API turns AI judgment into a building block you can actually budget and plan around.

When you are putting together a vibe-coded project that needs AI to make decisions (flagging user-submitted photos, routing a support message to the right flow, grading a form response against a rubric), you have usually had two options. Option one: prompt a general model and parse its text reply, which is slow and unpredictable. Option two: train a custom classifier, which requires data and engineering time most designers do not have.

The Decisions API is a third path. You define your options in plain language, pass in the content, and get back a structured answer with a confidence score. The confidence score is particularly useful: when the model is not sure, your app can route that case to a fallback or flag it for a human to review. That kind of graceful fallback is hard to build with open-ended text generation but straightforward here.

A concrete workflow to try

Google has a live playground in Google AI Studio where you can test the API without writing code first. That is a good first stop.

Once you are ready to wire it into a project, the practical shape of a request is: one shared input (a photo, a PDF, a voice note, whatever) plus a set of named questions. The input is billed once regardless of how many questions you attach, so batching is genuinely efficient rather than a marketing claim.

One real example from Google's own teams: Stitch, their UI design tool, uses the Decisions API in a rapid canvas experiment to generate playable wireframes from a described flow in seconds and to apply voice or tap edits almost instantly. The Decisions API is doing the routing and intent-classification work in the background, keeping the interactive loop feeling immediate. If you have used Stitch and wondered why the rapid mode feels snappier, this is likely part of the answer.

For your own builds, think about any moment where your app needs to sort, filter, rank, or gate content. Content moderation checks on uploads, routing a user's intent to the right next screen, scoring a submitted answer against a rubric, picking the right label from a fixed list. These are all decision shapes the API is designed for.

What to keep in mind

The Decisions API is built for closed-set choices. It is not for open-ended generation, summarisation, or anything where the answer space is undefined. If you need the model to invent an answer rather than pick one, you still need a standard generative endpoint.

The 256K token input limit covers very long documents and multi-modal combinations, but performance benchmarks in the announcement are based on contexts up to 16K tokens, and the API's latency stays nearly flat from 1 to 100 questions per request. How it behaves at the upper limits of both is worth testing before you rely on it in production.

There is also a no-cost tier in Google AI Studio for prototyping, which removes most of the risk in just trying it. At this stage, the most honest takeaway is: if you have a project where AI needs to make a lot of fast, structured judgments, this is the most purposeful tool Google has shipped for that job. It is worth an afternoon of experimentation before writing it off as too technical.

ai toolsgeminiapidesign toolsbuilding with ai

Sources

About VibeLab

VibeLab is a community for vibe coders. Discover vibe coded projects from around the globe, and post and host your own.