Google launches the Gemini Decisions API: a model that scores answers instead of writing them
Gemini Decision 1 returns probabilities for up to 1,000 closed questions per request at $0.10 per million input tokens, with no output charges. OpenAI launched a similar API in October.
Google on November 11 launched the Gemini Decisions API, a new Gemini API endpoint that doesn't generate text. It runs on a new model, Gemini Decision 1 (`gemini-decision-1`). You send one input and a set of closed questions, and the model returns a probability for each allowed answer instead of writing a response token by token. Google prices it at $0.10 per million input tokens and $0 for output, with a free tier in Google AI Studio for prototyping.
In plain terms, this is a classifier-and-judge endpoint for the small yes/no and pick-one calls that agent pipelines make constantly: which tool to call, which model should take a request, whether content breaks a policy, how good a retrieved passage is. Developers usually handle these today with a general model, a JSON schema and a prompt that says "answer only with one of these labels".
What Google announced
- Input: one shared input that can mix text, images, audio, video and PDFs, up to 256K tokens.
- Questions: up to 1,000 independent questions per request, of three kinds: yes/no probabilities, multiple choice across 2 to 255 options, and ordered rubric scores that return a weighted value plus per-level probabilities.
- Output: each answer comes with a confidence score from 0 to 1, which Google suggests using to send ambiguous cases to a fallback model or a human reviewer.
- Billing: 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 output or thinking-token charges and no separate fees for transcription or OCR.
- Latency: Google cites 50 to 100 ms of median model compute and 150 to 250 ms end to end, staying almost flat from 1 to 100 questions. That figure is for contexts up to 16K tokens, not the full 256K maximum.
- Availability: the Gemini API through Google AI Studio, with Python, JavaScript and REST guides. Google's code sample uses the `v1beta` API version. Google says support for Gemini in Google Cloud is coming soon, and that integrations are rolling out for OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode.
The claims that still need checking
Google says Gemini Decision 1 "leads across" benchmarks covering agent routing, reasoning, ranking and multimodal understanding, with the lowest calibration error. It also says the setup costs nearly 5x less across that benchmark suite than Gemini 3.5 Flash-Lite through the Interactions API. All of these numbers come from Google's own testing. The announcement doesn't name the competing models or publish the full suite, and Promptea found no independent evaluation. Read them as vendor claims for now.
Google also lists internal uses. Gemini Notebook routes queries to a model tier and thinking level, which Google says cut latency on everyday queries by up to 40%. Stitch uses it to edit wireframes quickly, CC from Google Labs uses it to pick emoji reactions, and Google Search uses it as a safety check on MCP tool use. These are Google's own descriptions, not outside case studies.
Not a one-off: the decision-model category
Google isn't first with this format. OpenAI released its own Decisions API in beta on October 6 with gpt-6-luna and made it generally available on October 22, raising the limit to 400 questions per request and adding prompt-cache discounts on the shared input. A day before Google's launch, on November 10, Hugging Face published a guide to "decision models" covering serving them with Transformers and llama.cpp and finding open models on the Hub. Its examples use the same state, questions and typed-answers structure.
Within about five weeks, two frontier labs and the main open-model hub have converged on the same idea: closed-set decisions deserve their own primitive, separate from text generation. Google's offer stands out on three points: native audio, video and PDF input, a higher question ceiling (1,000 versus OpenAI's 400) and no output charges. Promptea hasn't compared quality or real-world cost between the two, and neither vendor has published a direct head-to-head comparison.
What changes for developers
For anyone building agents or LLM pipelines, the practical change is where routing and grading logic lives. Many stacks today pay a generative model to write a label, parse the result, and retry when the format breaks. An endpoint that scores only predefined options can't return an invalid label, and it returns a probability that can drive thresholds directly. That replaces the usual "how confident are you, 1 to 10?" prompt, whose answers are poorly calibrated.
Prompting changes too. With this API, most of the work moves into writing the criteria for each option: clear, mutually exclusive descriptions of what each label means. The prompt engineering hasn't gone away. It now lives in the label definitions rather than in output-format instructions.
The limits matter just as much. These models don't explain their answers, and they can't handle open-ended questions. A wrong but confident probability is still wrong. Teams adopting the API should test calibration on their own data before trusting a threshold, especially at long context lengths, which fall outside the latency figures Google published.
Por que isso importa
- Routing, tool selection, policy checks and grading are some of the most frequent LLM calls in agent pipelines, and this API prices them by input only.
- It's the second frontier-lab API built specifically for closed-set decisions in five weeks, which points to a new, separate category of model endpoint.
- Calibrated probabilities can replace fragile JSON-label prompting and self-reported confidence scores.
Pontos principais
- Gemini Decision 1 handles up to 1,000 questions over one shared input of up to 256K tokens, mixing text, images, audio, video and PDFs.
- Pricing is $0.10 per million input tokens, nothing for output, plus a free tier in Google AI Studio.
- Google's 150-250 ms latency figure applies to contexts up to 16K tokens, and its benchmark and cost claims come from its own testing.
- OpenAI's Decisions API, generally available since October 22, allows up to 400 questions per request.
Fontes
- GooglePrimáriaGemini Decisions API: Fast, calibrated multimodal decisionsblog.google
- OpenAIPrimáriaOpenAI API changelog (Decisions API beta, Oct 6; general availability, Oct 22)developers.openai.com
- Hugging FacePrimáriaDecision Models in Hugging Facehuggingface.co
- gemini
- decision models
- model routing
- structured outputs
- api pricing
- coding agents
- Google DeepMind
- OpenAI
- Hugging Face
- Gemini Decision 1
- gpt-6-luna
- Gemini 3.5 Flash-Lite