Google published on 11 November 2026, under the byline of Min Choi (product manager, Google DeepMind) and Felix Yu (principal scientist), the Gemini Decisions API, powered by a new model: Gemini Decision 1 (gemini-decision-1). Coverage picked up on 12 November. The point is not to make an assistant write, but to make it decide: yes or no, one option among others, a score on a scale.

What the API actually does

You give it a shared input (text, image, audio, video or PDF, up to 256,000 tokens) and up to 1,000 independent questions. Instead of generating a reply word by word, the model scores the options in a single pass. Each answer comes with a probability and a confidence score (from 0 to 1). Google quotes 150 to 250 ms end to end, $0.10 per million input tokens, and no output tokens billed. The shared input is billed once, whatever the number of questions.

Three question types, mixable in the same request:

  • yes/no: the probability that a condition is true. In the API, Google calls this type noul;
  • choice: one option among 2 to 255, with probabilities for each candidate;
  • rating: an ordered scale, with a continuous score and probabilities per level.

The blog example pairs a photo of a countertop with a meeting note: is there a puddle, what to do, how serious. The official documentation, accessed on 12 November, stresses narrow questions, combined later in code, and using the confidence score to decide whether to automate or hand off.

Where Google already uses it

According to the blog, Gemini Notebook uses it to route a question to the right model: up to 40% lower latency on common queries, a publisher figure. Stitch uses it to iterate a product flow. CC, from Google Labs, picks an emoji among more than 200. Google Search uses it as a safety layer on MCP tools, to catch adversarial behaviour within the latency envelope. These are internal uses, described by Google.

The model is available through the Gemini API in Google AI Studio, with a free tier to prototype. Google cites integrations in progress (OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit, OpenCode) and upcoming support in Gemini on Google Cloud. InfoWorld, on 12 November, places the announcement a few weeks after a competing OpenAI service. The sources consulted do not name that service, nor mention a French-language interface.

What this changes for you

For an agent, choosing a tool, checking a rule or classifying a document does not need a paragraph. It needs a decision, a confidence threshold, and a path to a human when confidence is low. Before wiring such an API to a client email, an attachment or a business tool, it remains reasonable to fix the questions, the threshold, and what is not allowed to go out on its own.