AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Gemini Decisions API: Fast, calibrated multimodal decisions

Collected Nov 11, 2026

Google DeepMind introduced the Gemini Decisions API, a new endpoint built on a model called Gemini Decision 1 (gemini-decision-1), available through the Gemini API in Google AI Studio. The announcement is attributed to Min Choi, product manager at Google DeepMind, and Felix Yu, principal scientist. Developers can start today with an API key from Google AI Studio, using Python, JavaScript and REST guides.

The pitch is narrow but deliberate: instead of generating answer text token by token, the model scores a closed set of predefined options in one forward pass and returns probabilities and confidence scores. A single request carries one multimodal input, which can combine text, images, audio, video and PDFs, capped at 256K tokens, plus up to 1,000 independent questions. Google reports 150–250 ms end-to-end latency, with median model compute of 50–100 ms, and says latency stays nearly flat from 1 to 100 questions per request across contexts up to 16K tokens.

Pricing is $0.10 per 1M input tokens and zero for output tokens. The shared input is billed once per request no matter how many questions are asked; each question adds only its own instruction and criteria tokens. There are no variable thinking-token bills and no separate audio transcription or OCR charges. A no-cost tier in AI Studio is offered for prototyping. Google says that for a shared benchmark suite spanning agent routing, reasoning, ranking and multimodal understanding, this setup is nearly 5x cheaper than Gemini 3.5 Flash-Lite used with the Interactions API, and that Gemini Decision 1 leads the suite with the lowest calibration error.

The API exposes three question types that can be mixed in one request. The yes/no type estimates the probability that a condition holds, with optional descriptions of what counts as true and false. The multiple-choice type picks among 2 to 255 candidates and returns the winner plus normalized probabilities for every option. The ordered rubric type grades against a scale starting at 0 and returns a continuous weighted score plus per-level probabilities. Every answer carries a confidence score between 0 and 1, so calling code can send ambiguous cases to a fallback path or a human reviewer. Inputs use the same Content block format as the Interactions API, and resolution on image and video inputs is tunable to trade visual detail against token cost.

Several Google teams already use it as a fast control layer. Gemini Notebook routes incoming queries to the right model tier and thinking level, classifying intent in milliseconds and cutting latency up to 40% for everyday queries while saving deep reasoning models for hard ones. The rapid canvas experiment in Stitch generates playable wireframes of a product flow in seconds and applies voice or tap edits almost immediately, with one click handing screens to the full Stitch experience. CC from Google Labs picks an emoji reaction from a palette of more than 200 options in a fraction of a second. Google Search uses the API to detect adversarial behavior and content against adversarial evals within its latency budget.

This is a meaningfully different shape from a general chat model. Routing, content grading, policy checks and tool selection are classification problems, not writing problems, and applications usually only need a probability and a threshold rather than prose. Skipping autoregressive decoding removes the usual penalty where longer answers cost more and take longer, which explains why asking 100 questions costs little more in latency than asking one. In practice, the trade-off is that the model only answers within the options you define; whatever flexibility comes from open-ended generation is given up, and quality depends on how well the criteria are written. Calibration matters more than raw accuracy here, because a threshold of, say, 0.9 is only useful if a 0.9 score actually means roughly 90% likelihood. Google's claim of the lowest calibration error on its suite is the relevant number to watch, though independent checks on your own data would be the sensible next step.

Ecosystem integrations are rolling out with OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode alongside the google-genai SDKs, with support coming soon to Gemini in Google Cloud. Demo apps named Doorstep Parcels, Speed Sketch and Find the Zoo Keeper are available to try.

Why it matters: Product teams building agents, moderation layers and routers often burn a generative call just to make a small bounded choice. A purpose-built scoring endpoint priced per input token with no output billing could make those control-plane decisions cheap and fast enough to run on nearly every request. The main work shifts to defining good options and thresholds, and teams relying on confidence scores should validate calibration on their own traffic, since that is an inference from how the API is designed rather than a stated guarantee.

Read at Google AI Blog

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Meet the Gemini Decisions API, powered by our state-of-the-art model Gemini Decision for fast, calibrated multimodal decisions.