Google launches AI API for up to 1,000 fixed choices

Google introduced the Gemini Decisions API on Nov. 11, 2026, giving software a way to request fixed judgments from an AI model without receiving a written response. Applications can use the selected option and confidence score to route requests, check policies or rate content.

The service is available through the Gemini API in Google AI Studio and runs on the Gemini Decision 1 model. It supports yes-or-no questions, choices from a developer-provided list and ratings based on predefined scales.

A request can include up to 1,000 independent questions. Google reports end-to-end latency of about 150 to 250 milliseconds for requests containing one to 100 questions and contexts of up to 16,000 tokens. The reported latency measurements do not cover the maximum question or context limits.

Requests can combine text, images, audio, video and PDF files. The total context limit is 262,144 tokens across the shared input and all questions. Questions in the same request cannot access one another’s answers, so applications must combine the results after receiving them.

Google lists the price at $0.10 per million input tokens, with no separate charge for output tokens. Shared input is billed once per request, while each question adds its own instructions and evaluation criteria to the token count.

The API returns structured judgments that software can process directly. Google’s documentation distinguishes it from the Interactions API, which supports generated explanations, tool calls with arguments and multi-step reasoning.

Gemini Notebook uses the Decisions API to choose model tiers and thinking levels. Google reports up to 40% lower latency for everyday queries when the service reserves deeper reasoning for more complex requests. Search uses the API to detect adversarial behavior and content as a safety layer for its tools.

The confidence score is the model’s estimate that its selected answer is correct. Google recommends testing confidence thresholds against labeled examples from a developer’s own traffic and weighing the effects of incorrect approvals and rejections before an automated action is allowed.

A fixed list can produce a selected option even when none fits the input. Google recommends checking whether the input is within the question’s intended scope, adding an “other” option or using a separate validation step. Its documentation warns that user-provided text can attempt to influence its own classification.

Google recommends human review or a second check for high-stakes actions, including payouts. A confidence threshold does not guarantee that a decision is safe, and the model’s output does not determine what action an application should take.