Coding is Love
News 4 min read

Google launches Gemini Decisions API for fast yes/no and choice calls

Google released the Gemini Decisions API on November 11, 2026: a new Gemini API endpoint, powered by a model called Gemini Decision 1, that answers closed questions (yes/no, multiple choice or a score on a rubric) instead of writing text. It returns probabilities and a confidence score in about 150-250 ms, and it charges only for input tokens. For developers building agents, it is a cheap and fast way to handle routing, tool selection, policy checks and grading without asking a chat model to write out its answer.

What’s new

  • A model that scores, not writes. According to Google’s announcement, Gemini Decision 1 scores the candidate answers in a single forward pass rather than generating answer text token by token, and every answer comes with a confidence score from 0 to 1.
  • Three question types. A noul question gives the probability that a condition is true. A choice question picks from 2 to 255 options and returns a probability for each. A score question grades the input against an ordered rubric and returns a weighted score plus the probability of each level.
  • Many questions, one input. One request can carry up to 1,000 independent questions about the same input. The Decisions API docs put the context window at 262,144 tokens across the shared input and all questions.
  • Multimodal input. The input can mix text, images, audio, video and PDFs. Image and video resolution can be tuned to trade visual detail for token cost.
  • Speed. Google gives 150-250 ms end to end, with 50-100 ms of model compute, for contexts up to 16K tokens, and says latency stays nearly flat from 1 to 100 questions per request.
  • Pricing. $0.10 per million input tokens, and output tokens cost nothing. The shared input is billed once per request, however many questions you ask; each question adds only its own instruction tokens. There is a free tier in Google AI Studio for prototyping.

Google also says that on its benchmark suite (agent routing, reasoning, ranking and multimodal understanding) Gemini Decision 1 had the lowest calibration error, and that it was nearly 5x cheaper across that suite than Gemini 3.5 Flash-Lite with the Interactions API. The announcement doesn’t publish the individual scores in its text.

What it means for developers

The endpoint fits the small decisions inside an agent loop: which tool to call, which model to route a query to, whether a message breaks a policy, or how relevant a retrieved document is. Google says Gemini Notebook uses it as a real-time router and sees up to 40% lower latency on everyday queries, and Google Search uses it as a safety layer that checks MCP tool calls for adversarial behaviour.

It is available now in the Gemini Developer API (v1beta) and Google AI Studio, with guides for Python, JavaScript and REST. Support in Google Cloud is listed as coming soon. Google says integrations are rolling out in OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit and OpenCode; Vercel has added Gemini Decision 1 to its AI Gateway.

The docs say it needs version 3.2 or later of the google-genai Python SDK (pip install -U google-genai). Their quickstart looks like this:

from google import genai

client = genai.Client(http_options={"api_version": "v1beta"})

result = client.decisions.create(
    model="gemini-decision-1",
    input="The blue mug I ordered on Monday arrived cracked on Wednesday.",
    questions={
        "is_damaged": {
            "type": "noul",
            "instructions": "Did the item arrive damaged?",
        },
    },
)

answer = result.answers["is_damaged"]
print(f"P(yes)={answer.noul:.2f} confidence={answer.confidence:.2f}")

Some limits to know before you build on it. The docs list question instructions as text only, no YouTube links or gs:// Cloud Storage URIs as input, and no conversation history, thought or tool-call blocks. Context caching, the Batch API and Flex or Priority inference are not supported. Inline data can be up to 100 MB per request (50 MB for PDFs).

How it compares

Google is not first here. Simon Willison points out that the API is almost identical to Jev’s: the same three question types, and even Jev’s term “Noul” for a yes/no question. He writes that it costs the same as OpenAI’s GPT-6 Luna decision model, 10 cents per million input tokens with free output, while Jev is cheaper at 4.2 cents. The difference is input types: by his comparison, Jev takes text and JSON only and OpenAI’s model takes text and images, while Gemini Decision 1 also takes audio, video and PDFs.

He also noticed that Gemini gives less extreme probabilities on yes/no questions where the other two return 1.0 or 0.0. In his test with a short video of a robot, Gemini put “Does this video have a robot in it?” at about 0.92 and “Does this video have a dragon in it?” at about 0.05.

For Google’s newest general-purpose model, see our post on Gemini 4 Carbon.

Sources