Gemini Decision 1: Google's AI That Answers Instead of Chatting

Google launched a new kind of AI model on November 11, 2026 — one that never writes a word. Gemini Decision 1, the engine behind the new Gemini Decisions API, reads text, images, audio, video or PDFs and answers up to 1,000 yes-or-no, multiple-choice or rubric-scoring questions about them in a single pass, returning probabilities and confidence scores in about a quarter of a second. It costs $0.10 per million input tokens, with nothing charged for output, because there is no output text at all.

It is the clearest sign yet that decision models are becoming a category of their own, distinct from the chatbots that have dominated the AI boom. And for founders building on top of AI, that distinction matters more than the launch itself.

What a Decision Model Is

A chatbot generates text token by token. A decision model does something narrower and, for many applications, more useful: it scores a fixed set of options in a single forward pass over the input. Instead of asking a model to write an answer and then parsing the text, a developer sends one shared input — up to 256,000 tokens — along with a list of typed questions, and gets back a calibrated probability for every option.

The Gemini Decisions API defines three question types. A yes/no question checks whether a statement holds. A choice question picks one option from a list of 2 to 255. A score question rates the input against 2 to 255 ordered levels, like a rubric. Every answer comes with a confidence score from 0 to 1, so an application can hand unclear cases to a person or a bigger model.

Google says the model needs 50 to 100 milliseconds of compute and 150 to 250 milliseconds end to end, and stays nearly as fast whether it is asked one question or 100. It is available now through the Gemini API in Google AI Studio, with a free tier for prototyping; support for Vertex AI, Google's enterprise cloud platform, is described as coming soon.

Why Decision Models Are a New Category

The category did not start with Google. TypeSafe launched its Jev decision model in September, OpenAI unveiled its own Decisions API at its DevDay on September 29, and Microsoft followed with Microsoft-Decision-1, built on Alibaba's Qwen. Within weeks, all of the major AI labs had shipped a model that answers instead of writing. That kind of convergence usually means the market has found a real shape.

The shape is this: most of what software actually needs from AI is not prose. It is a judgement call. Is this transaction fraudulent? Which support ticket is urgent? Does this document comply with the policy? Those are closed-set decisions — the options are known in advance, and the value is in getting the answer right, quickly, cheaply and with a calibrated sense of confidence.

Chatbots can answer those questions, but wastefully. They generate tokens one at a time, their answers need parsing, their confidence is hard to measure, and their cost scales with the length of the answer. A decision model skips all of that. It is cheaper per decision, faster, and it returns a probability the calling code can branch on directly. At $0.10 per million input tokens with zero output tokens, a million decisions become a line item a startup can afford.

What It Means for Founders

For founders, the practical implications are immediate. Any product that today uses a chatbot to classify, route, triage or grade — and then parses the reply — can probably do the same job faster and cheaper with a decision model.

The deeper opportunity is architectural. Once decisions are cheap and calibrated, a product can make thousands of small judgements per user interaction instead of a handful. The constraint shifts from the cost of asking to the imagination of what to ask.

There are limits founders should respect. The model runs on the Gemini Developer API at API version v1beta only, with no Vertex AI listing yet, which matters for enterprises with strict cloud requirements. Google says free-tier data can be used to improve its products, which rules the free tier out for anything sensitive. And a probability is not a guarantee: calibration is only as good as the benchmark, and the benchmarks here are the vendor's own.

The larger point stands. The AI boom's first act was the chatbot. Its second act may be quieter — software that decides, at scale, for pennies — and the founders who build on that act early will have a cost and speed advantage that is hard to catch.

Angeled.com connects founders with investors.