Image: Google
Google has released the Gemini Decisions API and the model built for it, Gemini Decision 1. The model writes no text. It answers typed questions with probabilities that code can branch on directly. DeepMind product manager Min Choi and principal scientist Felix Yu announced it on Google's blog.
The API handles the small decisions a program makes constantly: which model should answer a request, which tool an agent needs, whether a message breaks a policy. With a chat model that means generating an answer and parsing it, and Google's pitch is to skip both steps.
Introducing Gemini Decision 1, our new frontier decision model in the Gemini API.
— Logan Kilpatrick (@OfficialLoganK) November 11, 2026
It is $0.10 per 1M input tokens with no additional cost. With support for image, audio, video, and PDF natively, 256K context window, & 1K questions.
You can try it for free in the API! https://t.co/Pa6pIaQyoy
Each request sends one shared input and a set of questions about it. Input can be text, an image, audio, video or a PDF. A yes or no question returns the probability that the condition is true, between 0.01 and 0.99. A choice question returns the winning option with a probability for every option. A score question rates the input against ordered rubric levels.
Every answer carries a confidence score, so code can act when confidence is high and hand the case to a person when it is not.
Google's example checks a photo of a breakroom counter with a short text note in one call. It asks whether there is an uncleaned spill, which cleanup action fits and how bad the mess is. Answers come back at 0.95 for the spill, "wipe with cloth" at 98% confidence, and a severity of 1.19 out of 2.
Image: Google
One request can carry up to 1,000 questions over shared input of up to 256,000 tokens. Questions are scored in parallel and cannot see each other, so extra questions add only their own instructions. Google says latency stays close to flat between one question and 100.
Input costs $0.10 per million tokens and output is free, because there are no output tokens. A no-cost tier covers prototyping in AI Studio. Google says the setup is nearly five times cheaper than Gemini 3.5 Flash-Lite with its text API across the shared benchmark suite.
Image: Google
Google's launch table puts Decision 1 ahead on seven of nine tests, including JevBench at 92.2% against 90.0% for TypeSafe Jev 1.14. It loses Mind2Web Multimodal at 86.2% to 86.8%, and When2Call at 79.9% to 82.0%. The stronger claim is calibration, where lower is better, at 0.0250 against 0.0494. That matters if the probabilities set thresholds. TypeSafe is cheaper on the shared evaluations, $0.68 against $0.88, and every number here comes from Google.
Question instructions can only be text, even when the shared input includes media. There is no Batch API, no context caching, no structured outputs, no function calling and no thinking mode. Media can come from a Files API reference, a public URL or base64 bytes, but not from YouTube links or Cloud Storage paths.
The docs warn that input is untrusted, since user text can argue for its own classification. The advice is to use explicit criteria, test adversarial examples and keep a human check on high stakes actions.
Google says Gemini Notebook uses the API to route queries to the right model tier, which it says cuts latency by up to 40% for everyday requests, and Search uses it to spot adversarial behavior in MCP tools.
It is a narrow product, and that is the point. Routing and policy checks need a number with a confidence attached, not a paragraph. It is on the Gemini Developer API in v1beta, with Cloud support promised later. We covered Google's Gemini 4 Carbon pricing and the work on why replies get faster without changing answers.
Comments