Skip to main content
株式会社オブライト
株式会社オブライト

What Is the Gemini Decisions API? A Decision-Only API That Scores Options in One Pass

Pricing, Usage, and How It Differs from Regular Gemini and Clef

Published 2026/11/1611 min read

The Gemini Decisions API is Google's decision-only endpoint (model gemini-decision-1, announced Nov 11, 2026) that scores a fixed set of options in a single pass and returns probabilities and confidence. Input is $0.10 per 1M tokens, output is free, and a free tier exists. Pricing, limits, code, and comparisons with regular Gemini, Clef, and Ollaya. Updated Nov 2026.

The Gemini Decisions API is a decision-only API from Google: given a fixed set of options, it scores them in a single pass without generating text and returns probabilities and a confidence score. The model ID is gemini-decision-1 (named Gemini Decision 1); only input tokens are billed, output tokens are always zero and free, and there is a free tier in Google AI Studio.

It was announced on November 11, 2026 on the official Google blog (by Min Choi and Felix Yu). It is offered on the Gemini API v1beta, with a Google Cloud version listed as coming soon. The announcement also collected 569 points on Hacker News (as of November 11, 2026). This article draws on the official blog and the Decisions API, pricing, and rate-limit pages in Google AI Studio (checked on November 16, 2026). Benchmark figures are Google's own claims and were not independently reproduced.

Why it is getting attention now

Agents and business automations make a huge number of pick-one decisions: which tool to call, whether a message violates a policy, which model to route to. Handing these to a regular LLM means generating the answer token by token, which is slower and makes the bill vary with output and thinking tokens. A decision model avoids that by scoring all options in one forward pass and returning only probabilities.

Since September 2026 there has been a run of decision models, from TypeSafe AI's Jev to Cloudflare Clef and the local runtime Ollaya. Google's entry matters because it is a first-party Gemini API feature priced at $0.10 per 1M input tokens with free output. The official blog cites Gemini Notebook's model routing (up to 40% lower latency on everyday queries) and Google Search using it as a safety layer for MCP tools that detects adversarial behavior.

What it can do

You provide one input (text, JSON, image, audio, video, or PDF) and send up to 1,000 typed questions about it in a single request. The shared input is read once, and each question returns probabilities. There are three question types.

Question typeWhat it doesWhat comes back
noul (yes/no)Estimates the probability that a condition holds; true_description / false_description are optionalnoul (0.01 to 0.99) and confidence
choice (multiple choice)Picks from 2 to 255 options, each with an ID and descriptionThe winner, probabilities for every option (summing to 1), and confidence
score (ordered rubric)Grades against ordered levels (2 to 255, 2 to 10 recommended)A probability-weighted continuous score, a legend, per-level probabilities, and confidence

Every answer carries a 0 to 1 confidence. The docs define it as the probability that the top answer is correct; for yes/no it equals max(noul, 1 - noul). The design assumes you route only low-confidence cases to a human or a stronger model. The blog lists tool selection, policy checks, model routing, and context grading as uses, plus multimodal pairings such as a call recording with account metadata or a scanned PDF claim with damage photos.

Pricing and getting started

ItemDetails
Model IDgemini-decision-1 (Gemini Developer API, v1beta only)
Paid price$0.10 per 1M input tokens (text, image, video, audio, PDF alike)
Output priceNone (output tokens are always 0)
Shared input billingBilled once per request; each added question only adds its own instruction and criteria tokens
Free tierThe pricing page lists the free tier as "Free of charge"; prototyping is possible with the API free tier in Google AI Studio
Input limit262,144-token context (shared input plus all questions)
Questions1 to 1,000 per request
Options2 to 255 for choice
Rate limitsFree: 200 RPM / 50,000 TPM / 10,000 RPD; Tier 1: 1,500 RPM / 1,000,000 TPM / 50,000 RPD; Tier 2: 5,000 RPM / 4,000,000 TPM / unlimited RPD; Tier 3: 15,000 RPM / 10,000,000 TPM / unlimited RPD
Latency (Google's figures)150 to 250 ms end to end, 50 to 100 ms median model compute (for inputs up to 16K tokens; roughly flat from 1 to 100 questions)
SDKgoogle-genai (Python) or @google/genai (JavaScript), version 3.2 or later
Not supportedContext caching, Batch API, Flex/Priority inference; conversation history, thought, and tool-call blocks as input
pip install -U google-genai
export GEMINI_API_KEY="(API key issued in Google AI Studio)"

The blog says this pricing is nearly 5x cheaper than Gemini 3.5 Flash-Lite with the Interactions API across the shared benchmark suite.

Quick start

- Create an API key in Google AI Studio (https://aistudio.google.com/api-keys) and set it as GEMINI_API_KEY
- Use SDK 3.2 or later and create the client with api_version: "v1beta"
- Pass the content to judge as input (the shared context) and a map of questions keyed by question ID as questions
- Read answers from answers under the same keys; route low-confidence cases to a person or a stronger model
- To just try it, there is an official playground (gemini-decisions-api-playground) in AI Studio

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?",
        },
        "department": {
            "type": "choice",
            "instructions": "Which department should handle this request?",
            "criteria": {
                "billing": "",
                "shipping": "",
                "technical": "",
                "other": "",
            },
        },
    },
)

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

c = result.answers["department"]
print(c)  # holds the winning choice, per-option probabilities and confidence (check attribute names in the official reference)

This code is assembled from the docs' one-question example (is_damaged) and the request spec (criteria is a map of option ID to description, and an empty description is allowed); it is not a verbatim official sample. Over REST, send the same JSON structure to POST https://generativelanguage.googleapis.com/v1beta/decisions with an x-goog-api-key header. The response contains answers and a usage object with input token counts (output tokens are 0).

Worked example — routing 10,000 inquiries a day

Assumptions: 10,000 inquiries a day over 30 days is 300,000 requests. Each request is assumed to be 500 input tokens, covering the inquiry text and the questions (a four-way department choice, an urgency yes/no, a three-level sentiment score). Output is free. Monthly input is 300,000 × 500 = 150M tokens.

ApproachCalculationApprox. monthly cost
Gemini Decision 1150M × $0.10 / 1MAbout $15 (output is free)
Regular Gemini 4 Hydrogen with JSON output (same 500 input tokens plus an assumed 50 output tokens)150M input × $1.50 + 15M output × $7.50About $337.5 ($225 + $112.5)

The 50 output tokens are a placeholder for a JSON answer; settings that emit thinking tokens would cost more. The gap comes mainly from not paying for output and from the lower input price. On rate limits, 10,000 a day sits exactly at the free tier's 10,000 RPD, so plan for a paid Tier 1 or higher in production. Real accuracy varies by use case, so validate against your own historical data before switching.

How it differs from alternatives

AspectGemini Decisions APIRegular Gemini (4 Hydrogen) with JSON outputCloudflare ClefOllaya / local decision modelsEmbeddings + similarity classification
MethodScores options in one forward passGenerates the answer token by tokenScores in parallel in one pass (open weights, Jev-compatible API)Open-source runtime that runs decision models locallyEmbeds text and classifies by distance
OutputProbabilities plus confidenceFree text / JSON (no probabilities)Calibrated probabilitiesProbabilities (model-dependent)Similarity scores
Price$0.10 per 1M input, output free$1.50 input / $7.50 output per 1M tokensModel is free (Apache 2.0); hosting cost is yoursRuntime is free (Apache 2.0), no token billing; hardware is yoursDepends on the embedding API (not researched here)
OperationsGoogle-managedGoogle-managedSelf-hosted (64k tokens, image support)Your own PC or serverSelf-hosted or vendor APIs
InputsText, image, audio, video, PDFSame (multimodal)Text and images (per official information)Model-dependentMostly text
Best forQuick start, multimodal judgments, high-volume routingDecisions that need an explanation or free-form textKeeping data in-house, retraining on your own dataFully local setupsMany fixed categories you maintain with example texts

As a rule of thumb: if you want a managed cloud API that is fast and cheap to start, or your judgments include images or audio, the Gemini Decisions API is a candidate. If you need an explanation or you generate text after deciding, regular Gemini 4 Hydrogen fits better. If data cannot leave your environment or you want to train on your own data, open-weight decision models such as Clef or Ollaya are the route. Embedding-based classification is light and easy to explain, but decision models are more flexible when you want to state the condition in plain language and get a probability.

Caveats and who it suits

- v1beta, Developer API only: it is available only on the Gemini Developer API v1beta, and Google Cloud support is listed as coming soon, so the spec may change
- No official word on Japanese: the docs do not describe supported languages, and Japanese accuracy has not been published, so test with your own Japanese data
- No caching or Batch: repeated input is billed in full every time, so estimate carefully for workloads with long shared context
- Confidence is not a guarantee: Google claims the lowest calibration error, but check calibration on your data before setting thresholds
- Question instructions are text only: put images and other media in the shared input
- Suits: developers who need to route inquiries, check policies, select agent tools, or route models at high volume. Does not suit: uses that need explained answers or long reasoning

FAQ

Is the Gemini Decisions API free to use?

The pricing page lists the Gemini Decision 1 free tier as "Free of charge", and you can prototype with the API free tier in Google AI Studio. The free-tier rate limits are 200 RPM, 50,000 TPM, and 10,000 RPD. The paid tier charges only input, at $0.10 per 1M tokens; output is always free.

How is it different from asking regular Gemini to pick via JSON?

Regular Gemini generates the answer token by token, while the Decisions API scores the options in a single forward pass and returns probabilities and confidence. There are no output tokens and no thinking-token variance in the bill. In exchange, it cannot give explanations or free-form text.

How many options and questions can I use?

Each request takes 1 to 1,000 questions. A choice question has 2 to 255 options, and a score question has 2 to 255 levels (2 to 10 recommended). The shared input plus all questions must fit in a 262,144-token context.

Does it work with Japanese inquiries?

The official docs and blog say nothing about supported languages, and there is no official information on Japanese accuracy. Check accuracy and confidence calibration on your own Japanese data first.

How should I choose between it, Cloudflare Clef, and Ollaya?

Pick Google's managed API for simplicity and low operational effort. If data cannot leave your environment, you want to retrain, or you need a fully local setup, open-weight Clef or the local runtime Ollaya fit better.

Can I use it on Google Cloud (Vertex AI)?

The blog lists Gemini in Google Cloud support as coming soon. As of November 16, 2026 it is available through the Gemini Developer API v1beta and Google AI Studio. Integrations such as OpenRouter, LangChain, and Vercel are described as rolling out.

Summary

The Gemini Decisions API is a decision-only API for handling pick-one judgments at volume for $0.10 per 1M input tokens with free output. Because it returns probabilities and confidence, it pairs well with a design that sends only uncertain cases to a person or a stronger model. Before adopting it, check that it is still v1beta, that Japanese support is not documented, and that caching and Batch are unavailable.

Choosing which model to put behind inquiry routing or an agent's decision step is something we also cover under AI consulting.

References (primary sources)

株式会社オブライト

Oflight Inc. is a Tokyo-based IT company providing software development, digital signage, web development, network systems, and IoT/AR/VR technology services.

Related free tools (no sign-up, instant results)

Feel free to contact us

Contact Us