Google has launched the Gemini Decisions API, a new endpoint that answers bounded questions instead of generating text. Powered by the new Gemini Decision 1 model, it scores up to 1,000 yes-or-no, multiple-choice or rubric questions against a single input in one pass, returning confidence scores in 150 to 250 milliseconds. It is available now in Google AI Studio at $0.10 per million input tokens, with no charge for output tokens.
A model that scores instead of writing
Gemini Decision 1 works differently from a normal language model. Rather than generating an answer token by token, it scores the candidate options in a single forward pass and returns probabilities and a confidence score for each question, not just a single answer.
That design keeps median model compute at 50 to 100 milliseconds, and Google says latency stays nearly flat from one question to a hundred at contexts up to 16,000 tokens. The point is scale: the same pass can score a hundred questions almost as cheaply as one, which is what makes the API viable as a control plane.
The API takes one multimodal input — text, images, audio, video or PDFs, up to 256,000 tokens — and evaluates up to 1,000 independent questions against it in a single request, sharing the input across every question, so the input is billed once no matter how many questions run.
The three question types cover yes-or-no checks, multiple choice from 2 to 255 options, and ordered rubric scoring. Every answer carries a confidence score from 0 to 1, so code can route ambiguous cases to a fallback or a human reviewer.
Google positions the endpoint for the small, numerous calls inside agent workflows: tool selection, policy checks, model routing and context grading. Teams inside Google are already using it that way. Gemini Notebook uses it as a real-time router, directing queries to the right model tier with up to 40% lower latency.
Google Search uses it as a safety layer for MCP tools, and the Stitch design tool uses it to generate playable wireframes in seconds. A no-cost tier is available in AI Studio for prototyping, with Google Cloud support listed as coming soon.
Our Read
Google is the third big lab to ship a decisions API, after OpenAI in September and Microsoft in October, and the pattern is now clear: the agent era needs cheap, fast judgment calls more than it needs more prose from a chatbot.
A routing decision that costs a full generation is a tax on every agent step; a scored decision at $0.10 per million input tokens with no output charge is close to free. That is why Google is selling this as a control plane, not a model.
The honest test will be calibration, not speed. Google says Gemini Decision 1 leads its benchmark suite with the lowest calibration error and is nearly five times cheaper than Gemini 3.5 Flash-Lite with the Interactions API across the same suite.
Those are company-reported figures with no independent testing yet, and a confidence score is only useful if it is right when it claims to be certain. Developers should test it on their own data before trusting it with routing decisions that matter.
Google is doubling the price of its flagship Gemini 4 Argon model from 1 December, which makes the Decisions API’s pricing look deliberate: the expensive generative models get dearer while the cheap judgment layer gets cheaper. Google is steering developers toward using the right model for each call, and charging accordingly.
Why It Matters
Agents make hundreds of small decisions for every visible answer — which tool to call, whether a result is safe, which model handles the next step. Each decision currently costs a generation, with its latency and its bill. A purpose-built scoring model collapses that cost, quietly reshaping what agents can afford to do.
The ecosystem is moving with it. Integrations are rolling out across OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode, alongside the google-genai SDKs. Google Cloud support is listed as coming soon, which is the remaining gap: until it lands, the API lives in AI Studio and the developer API rather than in enterprise cloud deployments.
Background
The Decisions API caps a busy run of Gemini releases. Gemini Omni 1.2 Flash reached general availability earlier this month, and Gemini gained a user-facing memory summary that lets people see, edit and delete what it remembers. Google is building Gemini as a platform of specialised models, and Decision 1 is the smallest, cheapest specialist yet.
Google published Python, JavaScript and REST guides with the launch, plus demo apps called Doorstep Parcels, Speed Sketch and Find the Zoo Keeper. The model identifier is gemini-decision-1, and it runs on the Gemini Developer API at v1beta, requiring the google-genai SDK 3.2 or later.
Source: Google’s announcement.
0 Comments