Generative models write text for people to read. When your app only needs to pick a tool or route a request, prompting a chat model for JSON adds latency and brittle parsing code you shouldn't have to maintain.
We built the Gemini Decisions API for these control-flow paths. Powered by our new Gemini Decision 1 model, it evaluates bounded questions in one forward pass. It generates 0 text tokens and returns calibrated probabilities in about 200 ms.
Each request takes one shared input across text, images, audio, video, or PDFs alongside up to 1,000 independent questions. Questions can't see each other and run in parallel, so latency stays nearly flat whether you ask 1 question or 100.
You can mix yes-or-no checks (noul), multiple-choice options (choice), and ordered rubric ratings (score) in a single call. Every answer includes a confidence score. High-confidence answers can trigger code paths right away, and uncertain cases can escalate to a reasoning model or a human.
At Google, Gemini Notebook already uses this endpoint to route prompts across model tiers, and Stitch uses it to apply voice edits to wireframes in real time.
You pay 10 cents per 1M input tokens (you only pay once for the shared context), and output tokens don't cost anything. Grab a free API key in Google AI Studio and read the docs at aistudio.google.com/docs/decisions-api.
