Google’s Gemini Decisions API: A Model That Won’t Write

Google launched the Gemini Decisions API in v1beta on November 11, powered by Gemini Decision 1, a discriminative model that returns calibrated probabilities for up to 1,000 yes/no, multiple-choice, or rubric questions per request and cannot generate text. Google prices Gemini Decision 1 at $0.10 per million input tokens, matching OpenAI's Decisions API, which launched October 22, caps at 400 questions, and does not accept audio or video; Google's API handles text, images, audio, video, and PDFs at 150–250ms end-to-end latency. Google says classification via a generative model such as Gemini 4 Hydrogen costs roughly 15 times more once output tokens are counted, though the beta label matters for production-critical work.

Google launched the Gemini Decisions API on November 11. The model powering it — Gemini Decision 1 — cannot write a sentence. It scores options. That is not a limitation; it is the design. Send it a document and up to 1,000 yes/no, multiple-choice, or rubric questions, and it returns calibrated probabilities for each. No output tokens. No prose. No parsing. Just numbers that tell you what the model thinks is true. What Gemini Decision 1 Actually Does Most classification pipelines today use a generative model: prompt it to pick option A, B, or C, then parse whatever it writes back. It works until it doesn’t — the model decides to explain itself, adds caveats, or returns something that breaks your regex. Gemini Decision 1 takes a different path. It is a discriminative model https://codelabs.developers.google.com/codelabs/discriminative-models , not a generative one. Rather than producing tokens one at a time, it scores all candidate answers in a single forward pass and returns probabilities. The output token count is literally zero. The model is architecturally incapable of generating text. The practical result: 150–250ms end-to-end latency for decisions across text, images, audio, video, and PDFs — all in a single request. Three Question Types Every request to the Decisions API shares a single context — your document, image, or audio clip — across a batch of questions. That shared input is billed once, regardless of how many questions you attach. Google supports three question types: - Yes/No noul : Returns a probability that a condition is met. Use it for policy checks, vulnerability detection, or relevance filtering. - Select choice : Pick from 2 to 255 options with a normalized probability per option. Use it for routing — which team, which tool, which model. - Rubric score : Rate on an ordered scale up to 255 levels. Use it for urgency triage or quality grading. A support ticket routing example using the Python SDK: python from google import genai client = genai.Client http options={"api version": "v1beta"} response = client.decisions.decide model="gemini-decision-1", input={"text": "Customer complaint: my order never arrived after 3 weeks"}, questions= {"type": "noul", "question": "Is this complaint urgent?"}, {"type": "choice", "question": "Route to which team?", "options": "billing", "shipping", "general support" }, {"type": "score", "question": "Urgency level", "max score": 5} Three routing decisions in one API call, against one shared input, with zero output tokens billed. Pricing and the OpenAI Comparison Google prices Gemini Decision 1 at $0.10 per million input tokens — the same number OpenAI charges for their own Decisions API https://munderdiffl.in/blog/gemini-decisions-api-vs-openai-decisions-api/ , which launched on October 22. The more relevant comparison is what you would pay using a generative model for the same task. Classification output from Gemini 4 Hydrogen is roughly 15 times more expensive once you account for the output tokens a generative model burns to explain itself before giving you the label you actually wanted. The two Decisions APIs differ on ceiling and modality. Google supports 1,000 questions per call and accepts audio and video. OpenAI tops out at 400 questions and does not accept audio or video. OpenAI’s version is generally available; Google’s is v1beta. That maturity gap is real — treat the beta label seriously if you are building something production-critical. When to Use It — and When Not To The Decisions API is the right tool when your output is a label, not a paragraph. Agent tool selection, content moderation checks, document routing, relevance scoring — these are all bounded decisions the API handles faster and more cheaply than a generative model. It is the wrong tool when you need extracted values, multi-field output, or any free text. The model cannot write. If a human or downstream system needs to read a sentence, use a generative model with structured outputs instead. One developer on DEV.to put it plainly: the Decisions API makes Gemini “a control plane, not a chatbot.” https://dev.to/siftsignal/googles-decisions-api-turns-gemini-into-a-control-plane-not-a-chatbot-396f This is infrastructure for agents — a fast, cheap decision layer that sits in front of your more expensive generative calls. Availability The Gemini Decisions API is live now in Google AI Studio and via the Gemini API v1beta https://blog.google/innovation-and-ai/technology/developers-tools/gemini-decisions-api/ . Google Cloud Vertex AI support is listed as coming soon. A free tier is available for prototyping. The required SDK is google-genai = 3.2.0 with the v1beta API version. If you are building multi-step agents with tool routing or content pipelines with classification steps, this is worth evaluating. The 1,000-question ceiling means you can run an entire document through a full rubric in a single call that costs fractions of a cent.