A lot of application code asks a model to choose between a short list of outcomes, then tries to parse the answer it writes. Google's Gemini Decisions API is designed for that narrower step. Announced on November 11, the endpoint uses Gemini Decision 1 to score yes/no checks, fixed choices or an ordered rubric, returning probabilities that application code can use directly. It is available through the Gemini API in Google AI Studio.
What does the Gemini Decisions API return?
A request contains an input and one or more questions about it. Each question is evaluated independently against the shared input and has a type: `noul` for yes/no, `choice` for a fixed set of options, or `score` for an ordered rubric. The response includes the selected answer or score, probabilities and a confidence value. Google says one request can include up to 1,000 questions, and accepts text, images, audio, video and PDFs, with input up to 256,000 tokens.
That shape fits work such as routing a support request, checking a document against a policy, or ranking an incoming item against a defined scale. The developer supplies the question and options, so the system is choosing inside a boundary the application sets. It is not an open-ended agent deciding what the app should do next.
How is it different from asking Gemini for text?
A generative call produces words that an application may need to interpret. A Decisions API call produces typed outcomes that code can branch on. Google says the endpoint does not generate output tokens and describes end-to-end latency of 150 to 250 milliseconds. Those latency and benchmark comparisons are Google's reported figures; results for an app will depend on its input, network, thresholds and fallback path.
The endpoint is a better fit when the possible outcomes can be listed in advance. A question like whether a receipt contains a damaged-item claim can be expressed as a yes/no check; picking one of several support queues is a choice; and assigning a severity against explicit levels is a rubric. If users need an explanation, a new idea, or an answer outside the list, a generative model or a human review step still has a role.
What does it cost, and is it available now?
Google lists Gemini Decision 1 at $0.10 per million input tokens and no charge for output tokens. The shared input is billed once per request, while each question adds its own instruction and criteria tokens. Google also offers an API no-cost tier for prototyping in AI Studio. Developers can call the endpoint with the Google GenAI SDK or REST; the published quickstarts specify the `v1beta` API.
The low listed token rate does not make every automated decision safe to run without review. Google's own documentation recommends setting confidence thresholds using labeled examples from the application's traffic and weighing false positives against false negatives. A confidence score is the model's estimate, not proof that a decision is correct. For account access, moderation, payments or other consequential actions, test against representative cases and route uncertain or high-impact outcomes to an appropriate human or independent check.
For a maker, the practical distinction is between generating a response and making a bounded choice that controls a workflow. The Decisions API gives code a structured answer for the latter. Whether it improves an app depends on how carefully the developer defines the choices, tests them and handles cases that do not fit.
Read Google's separate Gemini API and agent infrastructure update