Learn  /  Guides

Gemini Decisions API explained: yes/no, multiple-choice and scoring answers with probabilities, for 10 cents per million tokens

Most AI jobs in data work are not about writing text. They are about labelling: is this review a complaint, which team should get this ticket, how urgent is it? Google built a model just for that, and it returns probabilities you can actually set thresholds on.

Guide8 min readUpdated 16 November 2026
AIPythonData analysisWhat's new
All guides and cheat sheets

On 11 November 2026 Google launched the Gemini Decisions API, powered by a new model called Gemini Decision 1. It does not write answers. You give it some input and a list of questions with fixed answers, and it scores every option in one pass and returns the probabilities. For analysts, that is exactly what labelling, routing and quality checks need.

1,000Questions per callover one shared input
0.10Dollars per million tokensinput only, output is free
150Millisecondslowest end-to-end time quoted (up to 250)
255Optionsthe most in one multiple-choice question

Why a model that does not write

When you ask a normal chat model "is this a complaint?", it generates text, maybe adds an explanation, and you parse the answer. You pay for the output tokens, you wait for them, and "yes" tells you nothing about how sure the model was. Gemini Decision 1 reads the input once and scores each possible answer directly. Google quotes 150 to 250 milliseconds end to end, and says the time stays almost flat from 1 to 100 questions per request.

noul (yes/no)
The probability that a statement is true. You can describe what counts as true and what counts as false.
choice (multiple choice)
Picks one of 2 to 255 options and returns a probability for each.
score (rubric)
Grades the input on an ordered scale you describe, starting at 0, and returns a weighted score plus the probability of each level.
confidence
A number from 0 to 1 on every answer. Low confidence is your signal to ask a person.

The basic call

This is Google's own quickstart, with the google-genai Python package. Each question has a name you choose, a type and an instruction:

Python
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?",
        },
        "queue": {
            "type": "choice",
            "instructions": "Which support queue should handle this message?",
            "criteria": {
                "billing": "Payment, invoice, or refund questions.",
                "replacements": "Damaged, missing, or wrong items.",
                "general": "Product questions and account updates.",
            },
        },
    },
)

damaged = result.answers["is_damaged"]
queue = result.answers["queue"]
print(f"P(damaged)={damaged.noul:.2f} queue={queue.choice}")

Scoring on a scale

A rubric turns a fuzzy judgement into a number. You list the levels in order, from 0 upwards:

Python
result = client.decisions.create(
    model="gemini-decision-1",
    input="Export fails in Safari but works in Chrome.",
    questions={
        "severity": {
            "type": "score",
            "instructions": "How severe is this issue for the user?",
            "criteria": [
                "Cosmetic: appearance only, no lost functionality.",
                "Workaround available: a task fails, but another way works.",
                "Fully blocked: a task fails with no workaround.",
            ],
        },
    },
)

answer = result.answers["severity"]
print(answer.score, answer.confidence, answer.probabilities)

Labelling a whole column

Each call shares one input, so for a table you make one call per row and ask all your questions at once. A sketch with pandas, using the same calls as above:

Python
import pandas as pd

reviews = pd.read_csv("reviews.csv")          # columns: review_id, text

questions = {
    "is_complaint": {"type": "noul",
                     "instructions": "Is the customer complaining?"},
    "topic": {"type": "choice",
              "instructions": "What is the review mainly about?",
              "criteria": {"delivery": "Shipping, delays, damage in transit.",
                           "product": "Quality, size, features.",
                           "service": "Support, returns, staff.",
                           "price": "Cost, discounts, value."}},
}

rows = []
for r in reviews.itertuples():
    a = client.decisions.create(model="gemini-decision-1",
                                input=r.text, questions=questions).answers
    rows.append({"review_id": r.review_id,
                 "p_complaint": a["is_complaint"].noul,
                 "topic": a["topic"].choice,
                 "topic_confidence": a["topic"].confidence})

labels = pd.DataFrame(rows)
to_check = labels[labels["topic_confidence"] < 0.6]   # send these to a person

What it costs

You pay only for input: $0.10 per million tokens. The shared input is billed once per request, however many questions you ask, and each question adds only its own instruction and option text. A 100-word review is roughly 130 tokens; add about 100 tokens for two short questions and you have around 230 tokens per call. For 100,000 reviews that is about 23 million tokens, or roughly $2.30 at these prices. Check the current price list before a big run.

Decisions API

Labels with probabilities
  • No output tokens to pay for or wait for
  • Probabilities and confidence on every answer
  • Images, audio, video and PDFs in the same call
  • Only closed questions; it cannot extract free text or summarise
  • Runs outside your warehouse, so data leaves it

Best for routing, labelling and quality checks at volume.

AI functions in SQL

Labels inside the warehouse
  • Data stays in Snowflake, BigQuery or Databricks
  • Plain SQL, no code to host
  • You pay the warehouse's AI rates, usually per token in and out
  • Fewer probability details, depending on the platform

Best when the data is already in the warehouse and governance matters.

For the warehouse route, see AI functions in SQL: Snowflake, BigQuery and Databricks compared.

Where it fits, and where it does not

  1. 1

    Routing

    Send a question to the right team, queue or AI model. Google uses it to pick the model tier in Gemini Notebook, and says that cut everyday query time by up to 40%.

  2. 2

    Labelling for analysis

    Topic, sentiment, complaint or not, product mentioned: closed sets you can count and chart.

  3. 3

    Quality checks

    Grade AI answers or support replies against a rubric, then review the lowest scores.

  4. 4

    Policy checks

    Yes/no rules such as "does this message contain personal data?", with a human for the unsure ones.

  5. 5

    Not for

    Writing summaries, extracting free-text fields or answering open questions. Use a normal model for those.

Check your understanding

1/3What does Gemini Decision 1 return?

2/3How many questions can one call ask?

3/3What is a good use of the confidence score?

More: Claude Fable 5.2 vs Gemini 4 Carbon, embeddings and vector search for analysts and why AI agents need a semantic layer.

Sources: Google, "Gemini Decisions API: Fast, calibrated multimodal decisions" (11 November 2026); Google AI Studio documentation for the Decisions API and the Gemini Decision 1 model (code examples). The pandas example is our own sketch built on the documented calls. Speed, cost and accuracy figures are Google's.

Alessandro Ecclesie AgazziWritten by Alessandro Ecclesie Agazzi, analytics engineer in London. Updated 16 November 2026. About me

Found this useful?

New guides and "what's new" explainers on data and AI land here regularly. Follow along on LinkedIn, or share this one with a colleague.