Google opens Gemini Decisions API for fixed-choice judgments Google opens Gemini Decisions API for fixed-choice judgments

A smaller job than a chat model

The launch covers text, images, audio, video and PDFs in a shared input. Google reports 150–250 milliseconds of end-to-end latency for contexts up to 16K tokens. Its reported latency remains nearly flat between one and 100 questions. That measured range is narrower than the maximum context and question counts. Google describes uses inside Gemini Notebook, Stitch and Search, making this a component for deciding what happens next within a larger application.

The application owns the branches

The API documentation separates three operations: a binary judgment, a selection from named alternatives and a score over ordered levels. Questions evaluate independently against the same input. One answer cannot silently become another question’s evidence within that call. The application combines the results afterward. For a routing system, that means a scope check and a destination check can remain separate. A result supplies an option and probabilities, leaving ordinary program logic to determine which action follows.

Billing rewards shared inputs

Google’s price table charges the common input once per request. Each question still contributes its own instructions, criteria and formatting tokens. One million billed input tokens therefore costs ten cents; ten million costs one dollar. Those calculations exclude any later model call or action that the decision triggers. There is no output-token bill, while context caching and the Batch API are unavailable for this model. A free tier is listed alongside paid access. The table distinguishes their data use: free-tier content may improve Google’s products, while paid-tier content does not.

Structured answers solve a different problem

OpenAI’s August 2024 Structured Outputs release addressed a neighboring problem: making generated responses follow a developer’s JSON Schema. It uses constrained decoding, restricting which tokens can be emitted as the response is written. Its original explanation also distinguishes schema compliance from correctness inside the values. The comparison is architectural, not a speed contest between current models. A valid response format helps software consume an answer. A fixed-choice scoring interface gives software a different contract, centered on candidate alternatives and their probabilities.

The remaining work moves into evaluation

The practical consequence is a narrower unit to test. A team can measure its router separately from the model that eventually answers the customer. That separation makes a mistaken destination visible before it becomes mixed with the quality of a generated reply. It also makes the downstream action an explicit application decision, rather than something inferred from prose. A routing failure and a poor answer then become different defects with different owners and remedies.