Google has released the Gemini Decisions API, backed by a new model called Gemini Decision 1 (API name gemini-decision-1). It does exactly one thing: pick an answer from options you supply, and return a probability and a confidence score with it. It targets the small judgment calls that keep recurring in agent workflows, such as tool selection, policy checks, model routing and context triage.
In the official blog post, Google DeepMind product manager Min Choi and chief scientist Felix Yu explain that the model does not generate answer text token by token the way a typical large model does. Instead, it scores the candidate options directly.
“Gemini Decision 1 scores your candidate options in a single forward pass”
Three question types, up to a thousand at a time
The API supports three question types. Yes/no questions estimate the probability that a condition holds, and you can attach criteria for both the positive and negative case. Multiple-choice questions return the winning option plus a normalized probability for every candidate, with between 2 and 255 options allowed. Rating questions map to an ordered scale starting at 0 and return a weighted continuous score along with a probability for each level. Every answer carries a confidence value between 0 and 1.
Inputs can be text, images, audio, video and PDFs, and they can be mixed within the same request, for example a recorded customer-service call plus account metadata, or a scanned document plus a photo of damaged goods. A single request can take up to 256K tokens of input and evaluate up to 1,000 questions at once.
Speed is the headline of this release. Google gives an end-to-end latency of 150 to 250 milliseconds, with median model compute time of 50 to 100 milliseconds. In benchmarks with context of up to 16K tokens, latency stayed nearly flat as the number of questions rose from 1 to 100. Google's own Gemini Notebook already uses the model to route user queries to different model tiers in real time, cutting routing latency by up to 40%.
Input-only pricing, and a back-of-the-envelope calculation
Pricing is $0.10 per million input tokens, with no charge for output. In the blog's words, there are no output token fees, no bills that swell with the length of the model's reasoning, and no separate charge for audio transcription or OCR. Google says that on a shared benchmark, the approach is about 5 times cheaper than using Gemini 3.5 Flash-Lite through the Interactions API.
A rough illustration: an e-commerce support system sends in 10,000 tokens of call transcript and order data each time and asks 50 questions, such as whether to refund, whether the customer sounds angry, whether fraud is involved and which team should take the case. At the input price, that comes to about $0.001 per request, or roughly $1,000 per million requests. The estimate assumes the question text and the material together total 10,000 tokens; images and video are converted to tokens by resolution, so real figures will vary.
Be clear about what is being compared. Google benchmarked against small generative models such as Flash-Lite, where you pay for output and thinking tokens, so the bill grows with the number of questions and the length of the reasoning. The Decisions API removes that cost. The trade-off is that it can only answer closed-ended questions; open-ended summarization, rewriting and coding still belong with a generative model.
Until now, teams have mostly patched this need together in one of two ways: have a small generative model output a “yes” or “no” and then parse the result out of the text, or train a separate classifier that must be retrained whenever the labels change. The first gives no stable probability values, and the second cannot read images or audio. The Decisions API returns probabilities with confidence, so developers can set their own thresholds, for instance sending tickets with confidence below 0.7 to a human and letting those above 0.9 through automatically. How to pick those thresholds, and how closely the model's confidence tracks real accuracy, will have to be tested by developers on their own data; the blog only reports results from Google's internal use cases.
Where it is already used
Google lists several internal use cases: model routing in Gemini Notebook; fast canvas experiments in Stitch, where it modifies interactive wireframes in response to voice or click commands; CC from Google Labs, which uses it to pick a fitting reaction from more than 200 emoji; and Google Search, which uses it to add a safety check on top of MCP tools and detect adversarial behavior.
The API is available now through the Gemini API and Google AI Studio, with a free tier for prototyping and documentation covering Python, JavaScript and REST. Third-party integrations with OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, LiveKit and others will roll out over time. The Google Cloud version is marked “coming soon,” with no date announced.
Sources: Google's official blog post announcing the Gemini Decisions API (checked for latency, question cap, 256K input limit, per-million-input-token pricing and the Flash-Lite comparison basis), CocoLoop, public statements from the Google DeepMind product team.