Google launches Gemini Decisions API for fast, calibrated multimodal decisions

Google launches the Gemini Decisions API for fast, calibrated multimodal decisions

Google has introduced the Gemini Decisions API, a new endpoint powered by a new model called Gemini Decision 1 (model ID gemini-decision-1). It is available through the Gemini API and Google AI Studio. The announcement on the Google blog is credited to Min Choi, a product manager at Google DeepMind, and Felix Yu, a principal scientist there. The product does not write text. It answers closed-set questions: which tool to call, whether a policy rule is broken, which model should handle a request, or how well a retrieved passage fits a query.

What the API actually does

Google describes the target as bounded choices, such as tool selection, policy checks, model routing and context grading. The usual way to solve these is to ask a generative model for an answer and then parse the text. The Decisions API takes another path. Gemini Decision 1 scores the candidate options in a single forward pass instead of decoding token by token, and it returns probabilities and confidence scores. Google reports 150 to 250 milliseconds end to end, with 50 to 100 milliseconds of model compute at the median.

One request can carry a single multimodal input made of text, images, audio, video and PDFs, up to 256K tokens. Against that shared input you can ask up to 1,000 independent questions. The input field uses the same Content block format as the Interactions API, so teams that already use that API should find the shape familiar.

Three question types

Every answer comes with a confidence score between 0 and 1. Your code can use it to send unclear cases to a fallback path or to a human reviewer.

The first type is noul, a yes or no check. It estimates the probability that a condition holds, and you may add descriptions of what true and false mean. The second is choice, a multiple-choice question with 2 to 255 options. It returns the winner and a normalized probability for every candidate. The third is score, an ordered rubric that starts at level 0. It returns a continuous weighted score and the probability of each level.

Google's Python example checks a photo of a breakroom counter, together with a short text note, against all three types in one call. Is there an uncleaned spill? Which action fits: wipe with a cloth, call janitorial staff, or declare the room ready? How severe is the mess on a three-level scale?

Performance and price

Google says Gemini Decision 1 leads across a benchmark suite covering agent routing, reasoning, ranking and multimodal understanding, and that it has the lowest calibration error. Calibration matters here. A model that says it is 80 percent sure should be right about 80 percent of the time, and a system that routes on confidence depends on that.

On latency, the post says the median stays nearly flat from 1 to 100 questions per request, across multimodal inputs and contexts up to 16K tokens. On price, Gemini Decision 1 costs $0.10 per million input tokens, and output tokens are not billed. The shared input is billed once per request, and each question adds only its own instruction and criteria tokens. There are no variable thinking-token bills and no separate charges for audio transcription or OCR. Google adds that the combination is nearly 5 times cheaper across the shared benchmark suite than Gemini 3.5 Flash-Lite with the Interactions API. These are Google's own measurements and come with no independent verification evidence in this Google announcement.

How Google teams use it

The post lists four internal uses. Gemini Notebook uses the API as a real-time router that picks a model tier and thinking level for each query, which Google says cuts latency by up to 40 percent for everyday questions. The rapid canvas experiment in Stitch generates playable wireframes of a product flow in seconds and applies voice or tap edits almost at once. CC from Google Labs picks an emoji reaction from more than 200 options in a fraction of a second. Google Search uses the API as a safety layer for MCP tools, to detect adversarial behavior and content within its latency budget.

What it means for developers

The practical gain is that small decisions leave the generative path. Before, a team trained a custom classifier or paid for a full model call to get a one-word answer. Now you define options and criteria and read back probabilities. Because every answer carries a confidence value, engineers can set thresholds: act automatically when confidence is high, and hand off to a person when it is low.

There are limits. The comparisons come from Google's own benchmarks. The model only chooses among options you supply, so it will not replace tasks that need free text. And the post says support in Gemini on Google Cloud is coming soon, so enterprise customers must use the Gemini API for now.

Getting started

Get an API key in Google AI Studio, use the no-cost tier for prototypes, and follow the Python, JavaScript and REST guides in the documentation. Google also offers demo apps named Doorstep Parcels, Speed Sketch and Find the Zoo Keeper. Integrations are rolling out across OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode, alongside the google-genai SDKs.