Gemini API: Gemini Decision and the Decisions API for Calibrated Yes/No, Choice and Score Judgments
Gemini CLIGoogle released Gemini Decision (gemini-decision-1) and the new Decisions API in the Gemini API. The API returns calibrated probabilities for yes/no, multiple-choice and rubric-scored questions about text, images, audio, video and PDFs, and generates zero output tokens. Developers send up to 1,000 typed questions in one request and use the returned confidence values to decide whether to act automatically or escalate.
Key Takeaways
- Zero output tokens means Gemini Decision bills only for input, since the Decisions API never generates text.
- The API returns calibrated probabilities, so developers can set thresholds instead of parsing free-form model answers.
- Up to 1,000 independent questions can be asked about one input per request, within a 262,144-token window.
- Three question types cover most judgment work: yes/no, multiple choice and rubric scoring.
- Inputs can be text, images, audio, video or PDFs, so one model can moderate, route or grade mixed media.
- The
confidencefield supports automate-or-escalate workflows, although Google does not publish accuracy figures or the calibration method.
A Dedicated Model for Decisions, Not Text
Google added Gemini Decision (gemini-decision-1) to the Gemini API together with a new Decisions API, built for fast, calibrated binary, categorical and rubric decisions across text, images, audio, video and PDFs. Instead of generating a written answer that has to be parsed, it returns structured probabilities. Per the documentation, it generates zero text tokens, so output token usage is always 0.
How a Request Works
A request names the model (gemini-decision-1), supplies an input, and includes a questions map. The input can be a plain string, a single content block, or an array of blocks mixing text, image, audio, video and PDF content. Each question has a type and instructions, and choice and score questions also need criteria. Questions are scored independently, so one question cannot see the answers to the others. That lets a developer run many checks against the same input in a single call — from 1 to 1,000 questions, bounded by a 262,144-token context window shared across the input and all questions.
Three Question Types
- Yes/no questions return the probability of "yes". These probabilities always fall between 0.01 and 0.99, so they never claim certainty.
- Choice questions pick one option from a fixed list of 2 to 255 options. The response contains the top choice, a probability for every option, and a confidence value. Because probability is always spread across the options, a choice question still produces an answer even when none of the options fits well.
- Score questions rate the input against ordered rubric levels — 2 to 255 levels are allowed, with 2 to 10 recommended for most rubrics. The returned score is the probability-weighted average of the level indices, and a legend maps each index back to its description.
Working With the Output
The answers object reuses the keys sent in the request, and each answer matches the type of its question. The confidence field is the estimated probability that the top answer is correct. Google recommends using that confidence to choose between acting automatically and escalating to a person, with thresholds set from labeled examples. The documentation describes the probabilities as calibrated, but does not describe the calibration method or publish accuracy figures.
Limits and Availability
The Decisions API is available on the Gemini Developer API (v1beta) and works only with Gemini Decision 1. It does not support YouTube URLs or direct Cloud Storage (gs://) URIs, context caching, the Batch API, or Flex and Priority inference. Conversation history, thought blocks and tool-call blocks are not supported as input blocks, and question keys are limited to 1 to 64 ASCII characters matching [A-Za-z0-9_-]+. Only input tokens are billed.