Gemini Decisions API: Stop Generating Text When You Only Need a Yes or No

Gemini Decisions API: Stop Generating Text When You Only Need a Yes or No

Published
AuthorDevansh

On November 11, 2026, Google introduced the Gemini Decisions API, powered by a new model called Gemini Decision 1 (gemini-decision-1), available through the Gemini API in Google AI Studio. It answers a different question from a normal model. A normal model generates text, token by token, and you parse an answer out of it. The Decisions API scores a fixed set of candidate answers in a single forward pass and returns a probability and a confidence score for each one. No text is generated, so no output tokens are billed. That one design choice changes the cost, the latency, and the reliability of a whole class of agent jobs.

What it does

You send one multimodal input, text, images, audio, video, or a PDF, up to 256K tokens, plus up to 1,000 independent questions about it: yes/no questions, single-choice questions, and rubric-scoring questions. The model scores them all in one pass. Google quotes 150-250 ms end-to-end latency, with 50-100 ms of model compute, and says latency stays nearly flat from 1 to 100 questions per request. The intended uses are the bounded choices inside an agent: tool selection, policy checks, model routing, and context grading. Those are the decisions an agent makes dozens of times per task, and they are exactly the decisions you do not want to pay full generation prices for, or wait a full generation for, or parse out of free text that might say "yes, but" in three paragraphs.

The price is the point

Gemini Decision 1 is priced at $0.10 per million input tokens, with zero output tokens billed, and a no-cost tier in AI Studio for prototyping. Compare that with asking a full model the same question: you pay for the input and for every token it generates while it thinks out loud, and you pay in latency too. A router that costs a tenth of a cent per thousand decisions is a router you can afford to call on every step. A router that costs a full generation is a router you skip, and then your agent guesses. The honest limit is the same as the strength: it only answers closed questions. If the answer is not in your list of options, the model cannot invent a better one. Use it to choose, not to create. The creation still belongs to a full model.

Where it fits in a real agent

Three places, in order of value. First, routing: which model, which tool, which retrieval source for this step. A wrong route early wastes the whole task, so a cheap, calibrated router pays for itself immediately. Second, grading: is this retrieved context relevant, is this tool output trustworthy, is this draft good enough to ship or does it need another pass. A rubric score with a confidence number is a gate you can automate, where a free-text "looks good" is not. Third, policy: does this action need human approval, does this content cross a line you set, is this request in scope. A calibrated probability lets you set a threshold and log the borderline cases, instead of hoping the model remembers your rule. One practical note from API guides reporting on the launch: the Decisions API lives on its own endpoint, not the usual generateContent call, and SDK support is in google-genai 3.2 or later, so a model-ID swap will not work. Treat that as guide detail and check the official docs before you wire it in. If you are building your first agent and want the skeleton first, our n8n AI agent guide covers the loop, and our MCP explainer covers how tools reach the model. The Decisions API is the next layer: the cheap, fast judge that sits between the two.

What to watch before you trust it

Google emphasises calibration: the probabilities should mean what they say, so a 90 percent score should be right about 90 percent of the time. That is the claim to test on your own data, with your own questions, before you let a score auto-approve a destructive action. Start with routing and grading, where a wrong score costs a retry, not a production incident. Keep destructive actions behind human approval until your own logs prove the calibration holds on your workload. If you want the guardrails, the spend alerts, and the monitoring around an agent setup checked by someone who runs infrastructure for a living, book a free infra audit with DevOpsBY.

Disclaimer: This post is educational only, not financial advice. Figures are as reported by the sources on 2026-11-13 and may change - always confirm version numbers, requirements, and prices in the official docs before acting on them.

Related Articles

DevOps & Infrastructure Services

Need hands-on help? We build and manage the infrastructure behind the concepts covered in this article.

DevOpsBy Assistant

How can we help?

Select a topic to get started: