Summary
Google has introduced the Gemini Decisions API, which scores up to 1,000 questions against a shared input containing text, images, audio, video or PDFs. The company says it returns probabilities and confidence scores without generating answer text token by token.
Google introduced the Gemini Decisions API on November 11, 2026, as a way to evaluate many bounded questions against one shared input. Powered by the Gemini Decision 1 model, it accepts text, images, audio, video and PDFs and can return results for up to 1,000 questions in a request.
The API is intended for choices with a defined set of answers, such as selecting a tool, checking a policy or routing a request to a model. Google says the model scores candidate answers in one forward pass instead of generating a response token by token. It is available through the Gemini API in Google AI Studio.
How the decisions are scored
Developers can combine modalities in a single input—for example, an audio recording with account information—and ask several different kinds of questions about it. The API supports inputs of up to 256K tokens. Its three question formats cover yes-or-no checks, multiple-choice questions and ordered scoring rubrics.
For yes-or-no questions, the response estimates the probability that a condition holds. Multiple-choice questions can have 2 to 255 options and return normalized probabilities for each. A rubric score assigns a weighted result across ordered levels, starting at 0. Each answer includes a confidence score from 0 to 1, which software can use to send uncertain cases to a fallback or human reviewer.
A confidence score is an estimate, not a guarantee for an individual decision. Calibration describes how closely predicted confidence corresponds to observed correctness across many decisions. Google says Gemini Decision 1 had the lowest calibration error across its benchmark suite, which covered agent routing, reasoning, ranking and multimodal understanding.
Reported speed, cost and uses
Google reports median end-to-end latency of 150–250 milliseconds, including 50–100 milliseconds of model computation. It says latency stayed nearly flat as requests grew from 1 to 100 questions, for contexts up to 16K tokens. That latency description covers a smaller context than the API’s stated 256K-token input limit.
The announced price is $0.10 per 1 million input tokens, with no charge for output tokens; Google also offers an API no-cost tier for prototyping in Google AI Studio. The shared input is billed once per request, while each question adds its instructions and criteria tokens. Google says the API was nearly five times cheaper across its shared benchmark suite than Gemini 3.5 Flash-Lite with the Interactions API.
Google describes internal uses including routing queries in Gemini Notebook, selecting emoji reactions in CC and checking for adversarial behaviour in Search. These examples illustrate the API’s role as a fast decision layer: a system can define possible outcomes and use the model’s scores to choose what happens next. The company says developer guides are available for Python, JavaScript and REST, with integrations rolling out across several third-party developer platforms.