Google has released an API that never writes a sentence. The Gemini Decisions API, announced November 11 on Google’s blog, runs a new model called gemini-decision-1 that picks among options a developer defines and returns probabilities instead of prose.

That changes the price of every small judgment inside a product. Routing a query, checking a policy, grading a document, choosing a tool: each has a closed set of outcomes, and most teams have paid generative-model rates to settle them. Google charges $0.10 per million input tokens for this one and bills nothing for output.

Each request takes one shared input of up to 256K tokens, and that input can mix PDFs, video, audio, pictures, and text. Attached to it can be as many as 1,000 independent questions. Three question types exist. A yes/no question estimates the probability that a condition holds. A multiple-choice question selects from 2 to 255 options and returns a normalized probability for each. A rubric question grades the input on an ordered scale and returns a weighted score.

Every answer ships with a confidence value from 0 to 1. Google’s pitch is that low-confidence answers can be diverted in code, either to a backup path or to a person. That is the real product here: a classifier whose uncertainty is a number you can set a threshold on.

Google’s own example pairs an image of a messy breakroom counter with a note about a 9 AM client meeting. One call asks whether an uncleaned spill is present, which cleanup action fits (wipe with a cloth, dispatch janitorial staff, or declare the room ready), and how severe the mess is on a three-level rubric. Three answers come back from one pass over the image.

Speed is the second claim. Google says the model scores options without decoding token by token, which holds median compute to 50 to 100 milliseconds and end-to-end latency to 150 to 250 milliseconds. Latency barely moves whether a request carries one question or a hundred. Google also says the setup is nearly 5x cheaper than Gemini 3.5 Flash-Lite on the Interactions API across a shared benchmark suite.

Those comparisons are Google’s. The blog post claims the model ranks first across Google’s benchmarks and has the smallest calibration error, yet it publishes no accuracy figures and cites no outside evaluation. Calibration is also only as good as the traffic it was measured on, and a router or moderation gate sees a very different distribution than a benchmark does.

Google is already using the endpoint internally. Gemini Notebook routes each query to a model tier and thinking level with it, which Google says cuts latency by up to 40 percent on everyday requests. The Stitch design tool uses it to build playable wireframes in seconds. CC, from Google Labs, chooses among more than 200 emoji reactions. Google Search applies it as a safety layer on MCP tools to catch adversarial behavior.

The distribution plan matters as much as the model. Google says partner support is arriving through LiveKit, Pipecat, OpenCode, Vercel, LangChain, OpenRouter, LiteLLM, and Pydantic AI, with Gemini in Google Cloud listed as coming soon. Enterprise buyers on Google Cloud cannot adopt it there yet.

The target is the custom classifier and the cheap LLM call used as a gate. Both exist because teams needed fast, bounded answers and had no purpose-built option. A router that fires on every request is the cheapest place to find out whether a $0.10 decision model holds its confidence thresholds on your own traffic.

Reported by Google on its blog (blog.google) on November 11, 2026.