Google’s New Tool Doesn’t Write Answers. It Just Decides Yes or No.

A delivery photo lands in a company’s system. The software does not need a paragraph about the porch, the weather or the color of the box. It needs a few straight answers. Is the box damaged? Is it sitting under a porch? Should this go to a claims review?

That is the job Google is now selling. On Wednesday, CEO Sundar Pichai announced a new model built only to make those calls. It does not write a reply. It takes one batch of evidence, a photo, a PDF, a call recording, a video or plain text, and scores a checklist of up to 1,000 yes-or-no, multiple-choice or rating questions in a single pass.

Google calls the product the Gemini Decisions API, and the model behind it Gemini Decision 1. Product manager Min Choi and principal scientist Felix Yu laid out the details Wednesday on the company blog. Both work at Google DeepMind.

A checklist, not an essay

Most of the SI people use every day is built to talk. You ask a question and it writes an answer, word by word. Google’s argument is that a lot of real work is not a conversation. It is a decision.

You can mix question types in one request. A yes-or-no check returns the chance that something is true. A multiple-choice question can have as many as 255 options and comes back with a winner and a probability for each one. A rating question uses a scale the developer writes, from clean to a safety hazard, and returns a score. Every answer includes a confidence number from 0 to 1, so the software can send the close calls to a person.

The company’s own written example is a photo of a breakroom counter, plus a note about a 9 a.m. client meeting. The checklist asks whether there is an uncleaned spill, what to do next, and how bad the mess is.

Google’s public demonstration is even more ordinary. In a video the company posted, the model looks at 100 delivery photos and answers 16 questions on each one, including whether the box is damaged, whether it is under a porch, a 0-to-3 rating of the delivery, and where the case should go next. Google says that run finished in 4.5 seconds, and that it was nearly four times faster than the same job on its Gemini 3.5 Flash-Lite model. Those are Google’s numbers, from Google’s test.

Ten cents in, nothing out

Google is charging 10 cents per million tokens of input, the chunks the model reads, and nothing for the answer, because there is no written answer to bill. The shared evidence is billed once, even if you ask 1,000 questions about it. Each question adds only its own instructions. The company also has a no-cost tier in Google AI Studio for trying it out.

Google says a typical request comes back in 150 to 250 milliseconds, end to end, with 50 to 100 milliseconds of that spent on the model itself, when the input stays within about 16,000 tokens. It says that speed stays nearly flat from 1 question to 100. The overall cap is higher, about 256,000 tokens, enough for a long document, a set of photos or a recording, plus the questions.

On a benchmark suite Google ran itself, the company says the new setup is nearly five times cheaper than using Flash-Lite through its regular chat-style API. It also says the new model leads that suite and has the lowest calibration error, meaning its confidence scores are supposed to match how often it is actually right. No independent lab has published a review of those claims.

Google put it to work on its own products first

Google says its own teams are already using the tool as a fast control switch, instead of training a custom classifier or waiting on a long generated reply.

Gemini Notebook, the company says, uses it to send everyday questions to a cheaper, faster model and hold the heavier models for harder work, cutting wait time by up to 40% on those everyday queries. A new rapid-canvas experiment in Stitch, Google’s design tool, uses it to turn a product flow into playable wireframes in seconds and to apply edits made by voice or tap. A Google Labs app called CC uses it to pick an emoji reaction from a palette of more than 200 options. Search, Google says, is using it as a safety layer to flag hostile or evasive prompts inside the time those systems already have.

Developers can try three demos in Google AI Studio: Doorstep Parcels, Speed Sketch, and Find the Zoo Keeper. The company says connections are rolling out to OpenRouter, LangChain, Vercel and other developer tools, with support for Gemini in Google Cloud coming later.

The pitch is simple enough. A lot of software does not need a speech. It needs a yes, a no, and a number that tells you how sure the machine is, so a person can still catch the ones it should not decide alone.