Developers can use Google's Gemini Decisions API to apply up to 1,000 fixed-answer checks to one large submission, a format aimed at classification and routing tasks that do not require generated prose.
Google opened the API in AI Studio on Nov. 11. It runs on Gemini Decision 1 (gemini-decision-1). The shared input limit is 256,000 tokens, and supported formats span text, images, audio, video and PDFs. A video submission, for instance, can be evaluated against multiple yes-or-no tests together.
The API is built for tasks with a defined set of outcomes. It can return a classification, routing choice or other decision instead of open-ended text. Potential uses include selecting a tool for an AI agent, directing work to a model, checking policy compliance and evaluating retrieved context.
Google says confidence estimates are calibrated to how often decisions prove correct. At the 90% level, its aim is roughly nine correct decisions in 10. Google did not disclose a launch price.
Google's debut comes amid two other fall launches focused on machine decisions. OpenAI opened a public test of its Decisions API Oct. 6; Microsoft followed with Microsoft-Decision-1 on Oct. 9.
