Google Introduces Gemini Decisions API for Multimodal Choices

Google introduced the Gemini Decisions API on Nov. 11, giving developers a way to run fixed-answer decisions across multimodal inputs through its Gemini API.

The endpoint is available through Google AI Studio and runs on Gemini Decision 1 (gemini-decision-1). Developers can submit up to 256,000 tokens in one input, which may contain text, images, audio, video or PDF files.

The product is aimed at tasks with a limited set of possible answers. Instead of generating an essay, the model can return a decision such as which tool to select or whether a policy condition is met. A single request can include as many as 1,000 separate questions about the same input, allowing one file or recording to be checked against multiple criteria at once.

Google lists tool selection, model routing, policy checks and context grading as potential uses. Model routing directs work to a model, while context grading evaluates whether retrieved documents are relevant. Policy checks can classify material as allowed or flagged, and tool selection can determine an action for an AI agent. For moderation and compliance tasks, the checks can test content against rules. In retrieval systems, context grading can assess documents before they reach a main model.

Google describes the API's responses as fast and calibrated. Calibration means a confidence estimate should track how often an answer is correct: a 90% confidence level, for example, is intended to correspond to about nine correct answers out of 10. Google did not disclose pricing at launch.

The release follows two other decision-focused AI launches this fall. OpenAI's Decisions API public beta began Oct. 6, and Microsoft introduced Microsoft-Decision-1 on Oct. 9. Google's product arrived about five weeks after OpenAI's launch, adding another option for software that needs classifications, routing decisions or scores instead of open-ended text.

The Gemini API's published limits describe both the size of the shared input and the number of checks that can be grouped around it. For example, a video can be assessed against multiple yes-or-no criteria in one submission. The launch details list support for five input formats and up to 1,000 questions, but provide no price.