Gemini Decisions API: Fast, calibrated multimodal decisions

Open original

Google's new Decisions API lets developers get probability scores for closed-set choices like routing, policy checks, and grading, instead of generating text token by token. It's built on the Gemini Decision 1 model and aimed at fast, cheap control-plane decisions.

  • Gemini Decision 1 scores candidate options in a single forward pass, returning probabilities and confidence scores at 150-250 ms end-to-end latency.
  • Pricing is $0.10 per 1M input tokens with zero output tokens billed, and Google says it's nearly 5x cheaper than Gemini 3.5 Flash-Lite with the Interactions API on the shared benchmark suite.
  • Each request can carry one multimodal input (text, images, audio, video, PDFs up to 256K tokens) and up to 1,000 questions, with yes/no, multiple-choice, and rubric-score types that return confidence values so ambiguous cases can go to a fallback or human.

Google's new Gemini Decisions API is built for a narrow job: picking from a fixed set of options instead of writing free-form text. Think tool selection, policy checks, model routing, and grading context. The model, gemini-decision-1, doesn't generate answers token by token. It scores your candidate options in one forward pass and returns probabilities plus a confidence score for each. You can send a single multimodal input (text, images, audio, video, or PDFs, up to 256K tokens) along with up to 1,000 separate questions, and the claimed end-to-end latency is 150 to 250 ms. Pricing is $0.10 per million input tokens with zero output tokens billed, and Google says it comes out nearly five times cheaper than Gemini 3.5 Flash-Lite through the Interactions API on the shared benchmark suite. The API supports three question types. A "noul" question gives a yes/no probability. A "choice" question picks one of 2 to 255 options and returns normalized probabilities for all of them. A "score" question grades input against an ordered rubric and returns a weighted score. Every answer carries a 0 to 1 confidence value, which is the practical hook: your code can send low-confidence cases to a fallback model or a human instead of trusting a shaky label. The example in the post checks a breakroom counter photo with a text note against all three types at once, which shows how mixing modalities and question formats in one request is meant to work. Google also says latency stays nearly flat from 1 to 100 questions per request, which matters if you're batching checks. The examples Google gives are all internal or near-internal: Gemini Notebook routes queries to different model tiers and claims up to 40% lower latency on everyday requests, Stitch uses it to speed up wireframe generation, CC from Google Labs uses it to choose emoji reactions from over 200 options, and Google Search uses it as a safety layer for MCP tools. Those are vendor claims without independent numbers attached, so treat the 40% and the benchmark-leading calibration as things to verify yourself. The clearest takeaway for a developer is the trade-off: if your task is classification-shaped and you need an answer in a few hundred milliseconds with a usable confidence signal, this looks worth testing against your current setup. It is available now through Google AI Studio with a free tier for prototyping, and integrations with OpenRouter, LangChain, Vercel, LiteLLM, and several others are rolling out.

Ask about this article