Gemini Decision 1: Pricing and the Gemini Decisions API

Gemini Decision 1 pricing: what does zero output tokens mean?

Google's pricing page lists $0.10 per million input tokens across text, images, video, audio and PDF. It also lists a Free tier; actual access and quotas depend on your project. The Free tier is marked as used to improve Google's products, while the paid tier is marked No. Review the linked terms before sending production data.

The shared input is billed once per request, regardless of question count. Each question still adds tokens for its instructions, criteria and formatting. Therefore 1,000 questions are not free, and the advertised limit is not 1,000 unrelated documents or 1,000 sequential reasoning steps.

estimated_cost = total_input_tokens / 1,000,000 × $0.10

Illustrative total billable input	Google paid input cost	Interpretation
10,000 tokens	$0.001	Includes shared input and question overhead
20,000 tokens	$0.002	For example, 10,000 shared tokens plus 10,000 question tokens
1,000 requests at 20,000 tokens each	$2.00	Assumes the same token budget on every request

These are arithmetic estimates, not measured invoices. Use returned usage.total_input_tokens for the actual billable quantity. Zero output tokens means no generated prose to pay for; it does not mean the HTTP response is empty. The response contains typed answers, probabilities, confidence and usage. Context caching and Batch API are not supported in the checked model documentation.

For a generation workload, compare the Flash, Argon and Carbon cards separately. Their input/output rates and generated-token budgets are a different cost model.

How the Gemini Decisions API works

The model accepts text, images, video, audio and PDFs, with an input limit of 262,144 tokens. Questions operate independently over that shared state; one question cannot read another question's answer.

Question type	Request value	Returned result	Example use
Yes/no	noul	Probability of yes from 0.01 to 0.99, plus confidence	Is this package damaged?
Single choice	choice	Top option, per-option probabilities and confidence	Which support queue should receive this ticket?
Rubric rating	score	Probability-weighted level index, probabilities, legend and confidence	How severe is this issue?

Choice supports 2–255 options; score supports 2–255 ordered levels, with Google recommending 2–10 for most rubrics. Score levels start at index 0. A score of 1.4 is an expected index, not a claim that the answer belongs exactly to level 1 or a universal quality percentage.

For a dependent workflow, either ask both independent checks and combine their answers in code, or send a second request after the first decision. Google describes nearly flat latency as question count grows; that is not a guarantee for every input, network path or project quota.

Gemini Decision 1 Python quickstart

The official guide requires Google GenAI SDK 3.2 or later. Install the latest compatible package and set GEMINI_API_KEY to a Google-issued key:

pip install -U google-genai

import os
from google import genai

client = genai.Client(
    api_key=os.environ["GEMINI_API_KEY"],
    http_options={"api_version": "v1beta"},
)

result = client.decisions.create(
    model="gemini-decision-1",
    input="The mug arrived cracked. Please send a replacement.",
    questions={
        "is_damaged": {
            "type": "noul",
            "instructions": "Does the message report a damaged product?",
        },
        "queue": {
            "type": "choice",
            "instructions": "Which support queue should handle this message?",
            "criteria": {
                "billing": "Payments, invoices and refunds.",
                "replacements": "Damaged, missing or incorrect items.",
                "other": "Requests outside the listed categories.",
            },
        },
    },
)

answer = result.answers["queue"]
queue = answer.choice if answer.confidence >= 0.8 else "human_review"
print(queue, answer.probabilities)
print(result.answers["is_damaged"].noul)

This example follows the documented request schema; it was not executed as a paid model test for this article. The 0.8 threshold is illustrative. Select thresholds using labeled examples and the relative cost of false positives and false negatives.

An APIMaster key is not a Google API key. Do not send this request to an APIMaster chat endpoint or assume that replacing a model string adds native Decisions API support. Check the Decision model card for any later integration and its supported protocol.

Gemini Decision 1 vs Flash, Argon and Carbon

Choose Decision when your program needs a bounded judgment. Choose a generative Gemini model when it needs an explanation, a draft, code or a multi-step assistant response. A common architecture classifies a request first, then sends selected cases to a text model.

Requirement	Decision 1	Generative Gemini model
Typed probabilities over predefined options	Native Decisions API	A generated confidence number is not equivalent calibration
Free-form explanation or code	No text generation	Select a compatible generation route
Many independent checks on one document	Up to 1,000 questions per request	Prompt design and output budgets differ
Tool selection vs tool execution	Can choose among tool labels	Function calling depends on model and route
APIMaster availability at this check	No public active route	Flash 8; Argon 4; Carbon 4 public active routes

Decision's native typed responses are separate from the generative Gemini Structured Outputs feature. The model page lists text generation, thinking, function calling, search grounding and Live API as unsupported. It can select a tool label for your code, but does not itself execute that tool. Processing audio or video input also does not establish support for the Live API streaming protocol.

See the Carbon overview, Carbon vs Argon comparison and discounted Flash guide for the generative side of that workflow.

Confidence, evaluation and production thresholds

For yes/no questions, confidence is max(p_yes, 1 - p_yes). A confident No is therefore possible: a low yes probability does not imply low confidence. For choice and score, confidence tracks the most likely outcome, while a score's expected value can sit between levels.

Choice always distributes probability over the supplied options, even if none fits. Add an explicit other category or a separate scope check, and reserve a review path. Yes/no probabilities are bounded at 0.01–0.99, so a condition requiring a value above 0.99 never matches.

Before enabling automatic routing, evaluate held-out examples from your actual workload. Measure false positives, false negatives, calibration, review volume, latency and cost per accepted decision. Probabilities are model estimates, not proof that a policy check is correct. Keep high-impact actions behind appropriate application controls.

FAQ

Is gemini-decision-1 an official model ID?

Yes. Google's model documentation lists gemini-decision-1 for the Decisions API, with its latest update in November 2026.

How much does Gemini Decision 1 cost?

Google lists $0.10 per million input tokens on the paid tier. Shared input is counted once, each question adds its own input overhead, and there are zero output tokens. A Free tier is also listed, subject to project access and quotas.

Can I buy Gemini Decision 1 on APIMaster?

There were no public active routes at the 2026-11-13 check. The listed Google price is not an APIMaster sale price. Check the model card for current availability.

Does it answer 1,000 questions in sequence?

No. The limit is up to 1,000 independent questions over shared input per request. Questions cannot see each other's answers; dependent steps require orchestration in your code.

Can it replace a chatbot or coding model?

It does not generate text or code. Use it for typed classification, routing or scoring; use a compatible generative Gemini model for explanations and coding.

Does Gemini Decision 1 support Batch or caching?

The checked official model and pricing documentation list Batch API and context caching as unsupported. Sharing input across questions within one request is different from caching across requests.

Is its confidence guaranteed to be correct?

No. Validate probabilities and thresholds against labeled examples from your traffic. Include an out-of-scope option and a review path where mistakes are costly.

Get started with the right Gemini API

For native Decision 1, follow Google's Decisions API guide with a Google API key. For the generative stage, compare the Flash and Carbon cards, register with APIMaster, then create an API key. Set your compatible client's API base to https://apimaster.ai/v1 and choose an available model such as gemini-3.8-flash or gemini-4-carbon. The checked Carbon offer is up to ~82.1% below current Google Standard; the live card determines the price and supported features.

Sources

• Google: Gemini Decision 1 model specifications.
• Google: Decisions API guide and SDK examples.
• Google: Gemini API pricing.
• Google: Gemini API rate limits.
• APIMaster public marketplace API, queried by exact model ID on 2026-11-13; public enabled routes only.