Last updated: November 12, 2026
Gemini Decision 1 is a new Google DeepMind model, released November 11, 2026, that powers the Gemini Decisions API. It scores up to 1,000 yes/no, multiple-choice and rubric questions about one text, image, audio, video or PDF input in a single pass, returns calibrated probabilities, and costs $0.10 per million input tokens with no output charge.
Contents
- What is new
- Specs, pricing and limits
- How to make your first call
- The three question types
- Workflows and best practices
- How it compares
- Limitations
- Frequently asked questions
This guide is based on Google’s launch post, the official Decisions API documentation and the Gemini Decision model page in Google AI Studio.
What is new
Google says the new Decisions API is “a new endpoint for fast, closed-set decisions” built for tasks such as tool selection, policy checks, model routing and context grading.
Gemini Decision 1 does not write an answer. It scores your candidate options in one forward pass and returns a probability and a confidence value for each question. The documentation states that output token usage is always 0, so your code can branch on the numbers directly.
Google reports a median of 50–100 ms of model compute and 150–250 ms end to end, for contexts up to 16K tokens, and says latency stays nearly flat from 1 to 100 questions per request. It also says the API is nearly five times cheaper across its benchmark suite than Gemini 3.5 Flash-Lite with the Interactions API.
Google says Gemini Notebook already uses the API to route queries to the right model tier, serving everyday queries with up to 40% lower latency.
Specs, pricing and limits
| Item | Official value |
|---|---|
| Model code | gemini-decision-1 |
| Endpoint | Gemini Developer API, v1beta (Decisions API) |
| Inputs | Text, image, audio, video and PDF |
| Output | Calibrated probabilities; output token limit 0 |
| Context window | 262,144 tokens across shared input and all questions |
| Questions per request | 1 to 1,000 |
| Options per choice question | 2 to 255 |
| Levels per score question | 2 to 255 (2 to 10 recommended) |
| Inline data | Up to 100 MB per request (50 MB for PDFs); external URLs up to 15 MB each |
| Price | $0.10 per 1M input tokens; $0 for output; no-cost tier in AI Studio |
| SDK | google-genai 3.2 or later (Python or JavaScript) |
Figures come from the model page, the limits table and the launch post. For Google’s other model prices, see our Gemini API pricing answer.
How to make your first call
- Get an API key in Google AI Studio and set it as
GEMINI_API_KEY. - Install or update the SDK:
pip install -U google-genaiornpm install @google/genai@latest. Version 3.2 or later is required. - Create a client with
api_versionset tov1beta. - Call
client.decisions.createwithmodel="gemini-decision-1", one sharedinput, and aquestionsmap. Each question needs atypeand fullinstructions. - Read
result.answers[key]and act on the value only whenconfidenceclears a threshold you have tested.
A minimal routing call, adapted from the documentation:
from google import genai
client = genai.Client(http_options={"api_version": "v1beta"})
result = client.decisions.create(
model="gemini-decision-1",
input="I was charged twice for my order.",
questions={
"department": {
"type": "choice",
"instructions": "Which department should handle this request?",
"criteria": {
"billing": "Payments, invoices, and refunds.",
"other": "Anything outside these categories.",
},
},
},
)
answer = result.answers["department"]
print(answer.choice, answer.confidence)For a terminal workflow, see our Gemini API CLI guide.
The three question types
| Type | Use it to | Returns |
|---|---|---|
noul (yes/no) |
Check whether a statement holds | Probability of yes (0.01 to 0.99) and confidence |
choice |
Pick one option from a fixed list | Top option, per-option probabilities, confidence |
score |
Rate input on ordered rubric levels | Weighted score, per-level probabilities, legend, confidence |
Three details trip up new users. Yes/no probabilities never exceed 0.99, so a threshold above 0.99 never matches. Choice keys are sorted alphabetically before evaluation, so use a catch-all key such as other rather than none_of_the_above. And the score is a probability-weighted average of level indices, so check the per-level probabilities when an input could split between levels.
Workflows and best practices
Google’s documentation lists routing, multi-rule policy checks, agent tool selection, rubric ranking and real-time reactions to UI state, audio or video as good fits.
Practical patterns from the official best practices:
- Ask atomic questions. Keep one decision per question and combine answers with ordinary
ifstatements in code. - Write the full question. The model never sees question keys or variable names.
- Batch checks on one input. The shared input is billed once, so ten policy checks cost little more than one.
- Do math in code. Compare numbers, dates and durations yourself, then pass the result in the input.
- Treat input as untrusted. User text can argue for its own classification, so keep a human or second check on payouts and other high-stakes actions.
- Control media cost. Set image and video resolution to low, medium, high or ultra_high (images only for ultra_high).
For agent builders, a common setup is to use Gemini Decision 1 as a cheap router in front of larger generative models, similar to the routers in our OpenRouter Auto Code guide.
How it compares
Decision models are now a small category. Simon Willison notes that Google’s API is “almost identical to Jev” and even reuses Jev’s “noul” term for yes/no questions.
| Model | Input price per 1M tokens | Input types |
|---|---|---|
| Gemini Decision 1 (Google) | $0.10, output free | Text, images, audio, video, PDF |
| GPT-6 Luna decision model (OpenAI) | $0.10, output free | Text and images |
| Jev | $0.042 | Text and JSON |
The comparison figures for OpenAI and Jev are as reported by Simon Willison, not taken from those vendors’ pages. Willison also found that Gemini returns less extreme yes/no probabilities than the other two, which return 1.0 or 0.0. For other decision APIs, see our guides to the Perplexity Decisions API and Hindsight decision models.
Within Gemini, use the Interactions API instead when you need generated text, extracted values in your own JSON schema, tool calls or multi-step reasoning.
Limitations
- Gemini Developer API (v1beta) only; Google says Gemini in Google Cloud support is coming soon.
- No context caching, Batch API, or Flex or Priority inference. Repeated input across requests is billed in full.
- No YouTube URLs or direct
gs://URIs; use the Files API, HTTPS URLs or base64 data. - Question instructions accept text only, and conversation history, thought and tool-call blocks are not supported as input.
- Questions are independent. Dependent decisions need a second request or code-side logic.
- Requests blocked by safety filters return HTTP 400 with a code such as
safetyorprohibited_content.
Frequently asked questions
What is Gemini Decision 1?
Gemini Decision 1 (model code gemini-decision-1) is a Google DeepMind model released on November 11, 2026 for closed-set decisions. It answers yes/no, multiple-choice and rubric-score questions about text, images, audio, video and PDFs, and returns calibrated probabilities instead of generated text.
How much does the Gemini Decisions API cost?
Google prices Gemini Decision 1 at $0.10 per 1 million input tokens and $0 for output tokens. The shared input is billed once per request no matter how many questions you ask. There is also a no-cost API tier in Google AI Studio for prototyping.
How many questions can one Decisions API request include?
One request can include 1 to 1,000 questions, all evaluated against the same shared input. The total of the input and all questions must fit in the 262,144-token context window. Each question is scored independently, so questions cannot see one another.
Can Gemini Decision 1 generate text or call tools?
No. Google’s model page lists text generation, function calling, structured outputs, thinking, search grounding and the Live API as not supported. Use the Interactions API with a generative Gemini model when you need written answers, extracted values or tool calls.
How does Gemini Decision 1 compare with Jev and OpenAI’s decision model?
According to Simon Willison, all three support yes/no, choice and rubric questions. Gemini Decision 1 matches GPT-6 Luna at $0.10 per million input tokens, while Jev costs 4.2 cents. Jev is text only, OpenAI handles text and images, and Gemini adds audio, video and PDFs.
Sources
- Gemini Decisions API: Fast, calibrated multimodal decisions (Google, November 11, 2026)
- Decisions API documentation (Google AI Studio)
- Gemini Decision model page (Google AI Studio)
- Gemini Decisions API (Simon Willison, November 11, 2026)