On November 11, 2026, Google announced the Gemini Decisions API on its official blog, powered by the new Gemini Decision 1 model. It does not write articles or chat. It does one thing: make closed-set judgments about a given input. A single request can pack in up to 1,000 yes/no, multiple-choice, or scoring questions, and the model answers them all in one forward pass, with end-to-end latency of 150 to 250 milliseconds. For products that need decisions in real time, that is far faster than waiting for a generative model to produce its answer token by token.
What one input can judge
The API accepts mixed input — text, images, audio, video, and PDFs — up to 256K tokens per request. Three question types can be mixed in a single request: yes/no checks return the probability of yes or no; multiple-choice questions support 2 to 255 options and return the winner plus normalized probabilities for every candidate; scoring questions use an ordered rubric starting at 0 and return a continuous score with per-level probabilities. Every answer carries a confidence score from 0 to 1, so code can route uncertain cases to a human or a fallback flow. Pricing is $0.10 per million input tokens with output tokens free, and the same input is billed once no matter how many questions hang off it. Google's published benchmarks show Gemini Decision 1 leading on agent routing, reasoning, ranking, and multimodal understanding, with the lowest calibration error — running the full suite costs nearly 5x less than Gemini 3.5 Flash-Lite with the Interactions API.
Who will use it first
Google already has four deployments inside the company: Gemini Notebook uses it for real-time routing, sending each query to the right model tier and thinking level and cutting everyday query latency by up to 40%; Stitch's rapid canvas uses it to generate playable product-flow wireframes in seconds; CC from Google Labs uses it to pick the right emoji from a palette of more than 200 in an instant; and Google Search uses it as a safety layer for MCP tools, detecting adversarial behavior and content. The fitting scenarios are specific: tool selection, policy checks, model routing, and context grading. The boundary is equally clear: it does not generate open-ended text, so tasks that need long answers, creative writing, or complex reasoning still belong to generative models. Ecosystem integrations are rolling out: OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit, and OpenCode are already on board, with support in Google Cloud still being prepared.
The API is already available to try in Google AI Studio, with a no-cost tier, and Google has released three demo apps — Doorstep Parcels, Speed Sketch, and Find the Zoo Keeper. The thing to watch is whether its calibration holds up on real business data as well as it did in benchmarks — once confidence scores go wrong, the cost in routing and moderation scenarios grows immediately.
