The Gemini Decisions API is a new Google endpoint that answers fixed questions about text, images, audio, video and PDFs with probabilities instead of generated text. It runs on Gemini Decision 1. Teams building agents should use it for multimodal checks that sit before an action, such as routing, policy and evidence checks, and keep generation elsewhere.
What was released
The Gemini Decisions API was announced by Google on 11 November 2026 in the post Gemini Decisions API: Fast, calibrated multimodal decisions. It is powered by a new model, gemini-decision-1, and is available through the Gemini API in Google AI Studio. The code sample in the announcement calls client.decisions.create from the google-genai Python SDK with the v1beta API version, so treat the interface as one that may still change.
Google describes it as a set of primitives for bounded choices: tool selection, policy checks, model routing and context grading. Rather than writing an answer token by token, the model scores the options you supply in a single forward pass.
What it does
One shared input, many questions. A request carries one input of up to 256K tokens and up to 1,000 independent questions about it. The input can mix text, images, audio, video and PDFs, for example a call recording with account metadata, or a scanned claim form with damage photos. Resolution on image and video inputs can be tuned to trade visual detail against token cost. The input uses the same content block format as Google's Interactions API.
Three question types. A yes/no question returns the probability that a condition holds, with optional criteria describing what counts as true and false. A choice question picks from 2 to 255 options and returns the winner plus a probability for each option. A score question grades the input against an ordered rubric starting at level 0 and returns a weighted score with a probability for each level. Every answer carries a confidence value from 0 to 1.
Latency. Google gives 150-250 ms end to end, with median model compute of 50-100 ms, staying nearly flat from 1 to 100 questions per request for contexts up to 16K tokens.
Billing. Gemini Decision 1 has a per-token input price and no charge for output tokens. The shared input is billed once per request however many questions you ask; each question adds only its own instruction and criteria tokens. There are no thinking-token charges and no separate audio transcription or OCR charges. A no-cost tier in Google AI Studio is available for prototyping.
Where it runs. Today that is the Gemini API through Google AI Studio and the google-genai SDKs. Google says integrations are rolling out across OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode, and that support for Gemini in Google Cloud is "coming soon".
How it differs from the other Decisions APIs
OpenAI and Perplexity both shipped endpoints with the same name this autumn, and CodeDTX has covered OpenAI's general availability release and Perplexity's version. The idea is shared: typed answers with probabilities, no prose. The practical differences for Gemini's version, going by Google's announcement, are:
- Audio, video and PDFs go straight in. You do not need a transcription or OCR step before asking a question about a recording or a scanned document, and you are not billed for one.
- The question limit is higher, at 1,000 per request against one shared input.
- The shared input is billed once, which matters when many questions read the same long document or recording.
- Enterprise hosting is not there yet. Until Gemini in Google Cloud supports it, the endpoint sits under the Gemini API terms rather than your Google Cloud agreement.
The request and answer formats differ between all three, so do not reuse code across them. Put the decision call behind your own interface so that the provider is a configuration choice.
Which agent checks to move to it
CodeDTX's view: the strongest fit is the set of checks an agent runs before it acts, where the evidence is not plain text.
Evidence checks on files and media. An agent that approves a claim, a return or a delivery query often has a photo, a scan or a recording as its evidence. Today that usually means a transcription or vision call, then a text prompt, then a parser. One request with yes/no questions such as "does the photo show damage to the item?" and "does the form carry a signature?" replaces that chain, and the probabilities give your code something to compare with a threshold.
Routing before generation. Google says Gemini Notebook uses the endpoint to send each query to a model tier and thinking level. The same pattern works for any agent that has a cheap path and an expensive one: a choice question decides which model, tool or queue handles the request, and generation only starts once that is settled.
Policy and tool-safety checks. Google also names a safety layer for MCP tools in Search. Asking several policy questions about a proposed tool call in one request, against the same input, keeps the check inside a latency budget. It does not replace permission controls; it adds a fast screen in front of them, alongside the checks you run when you vet third-party MCP servers.
Context grading. Score questions can rate each retrieved passage or attachment for relevance before it goes into the main model's context, which keeps long contexts smaller.
Leave these where they are. Anything that must produce fields, an explanation or a tool call with arguments needs a generative model. Questions whose answers depend on each other need separate requests, because the questions in one request are independent.
What to set up before you switch
Thresholds with an owner. A confidence of 0.7 means nothing until your code decides what happens above and below it. Set thresholds from labelled examples of your own traffic, decide which band goes to a person, and record who reviews them, as you would for any confidence score that sends work to human review.
A side-by-side test. Run the endpoint and your current classifier prompt or pipeline over the same labelled cases. Google reports benchmark results for calibration, but your categories, file types and edge cases are not in its suite.
Token cost of media. Billing is input only, so the cost of a check is driven by how much image, video and audio you send. Set resolution deliberately and cut recordings to the part that matters.
Data terms. Because the endpoint runs under the Gemini API rather than Google Cloud for now, check how request data is handled under those terms before you send customer recordings or documents, and compare it with what your contracts allow. The same review applied to Gemini API request storage is a sensible starting point.
When not to switch
Wait if your workload must run under a Google Cloud agreement, in a specific region, or under contract terms that the Gemini API does not cover; Google says that support is coming but gives no date. Stay with a small self-hosted classifier if it already meets your accuracy, latency and data rules. Keep a generative call where an audit trail needs the model's reasoning written down. Move when an agent's pre-action checks currently chain transcription, OCR or vision calls into a text prompt and a parser.
Frequently asked questions
What is the Gemini Decisions API?
It is a Google endpoint, announced on 11 November 2026, that answers fixed questions about an input instead of generating text. It runs on the Gemini Decision 1 model through the Gemini API in Google AI Studio. You send one input of text, images, audio, video or PDFs and up to 1,000 yes/no, multiple choice or rubric questions, and each answer comes back with probabilities and a confidence score.
How is the Gemini Decisions API billed?
Google charges per input token for Gemini Decision 1 and does not charge for output tokens. The shared input is billed once per request regardless of how many questions you ask, and each question adds only its own instruction and criteria tokens. There are no thinking-token, audio transcription or OCR charges. A no-cost tier in Google AI Studio covers prototyping. Check the announcement for the current rate before estimating costs.
How is it different from the OpenAI Decisions API?
Both return typed answers with probabilities rather than prose. According to Google, its version accepts audio, video and PDFs directly, allows up to 1,000 questions per request and bills the shared input once. It currently runs through the Gemini API in Google AI Studio, with Google Cloud support announced as coming soon. The request and answer formats differ, so code written for one provider will not work unchanged on the other.
Can I use the Gemini Decisions API on Google Cloud?
Not yet, according to the announcement. Google says support for Gemini in Google Cloud is coming soon but gives no date, and links a sign-up form for interest. Until then the endpoint is reached through the Gemini API with an AI Studio key. If your organisation needs the data and contract terms of a Google Cloud agreement, keep production workloads on your current approach and prototype in AI Studio.



