Overview
Gemini Decision 1 is Google's first decision model, released on November 11, 2026, alongside a new Gemini Decisions API available through the Gemini API and Google AI Studio under the model ID gemini-decision-1.
It works differently from a chat model. Rather than asking for a paragraph, you give it one multimodal input and a batch of questions that must have definite answers, and it returns a probability and a confidence score for each. Questions come in three kinds and can be mixed in a single request. Noul is yes or no, estimating the probability that a condition holds and optionally carrying separate criteria for true and false. Choice picks from 2 to 255 options and returns the winner plus normalised probabilities for every candidate. Score rates against an ordered rubric starting at zero and returns a continuous weighted score with per-band probabilities. One request takes up to 1,000 independent questions, the input cap is 256K tokens, and text, images, audio, video and PDFs can all go in directly.
Latency is the headline. Google quotes 150 to 250 ms end to end, with median model compute of 50 to 100 ms, measured on contexts up to 16K tokens. The more interesting property is that question count barely moves it: going from 1 question to 100 leaves latency nearly flat, because the judgements resolve in a single forward pass instead of expanding into separate generation tasks.
The price structure is the point of the design. Input costs $0.10 per million tokens and output is free. The shared input is billed once per request regardless of how many questions ride along, and each question only adds the tokens of its own instruction and criteria. Google states there are no output token charges, no variable thinking token charges, and no separate audio transcription or OCR fees. A free tier in AI Studio covers prototyping.
On capability, Google says the model leads across a shared benchmark suite spanning agent routing, reasoning, ranking and multimodal understanding, with the lowest calibration error, and that against Gemini 3.5 Flash-Lite with the Interactions API it comes out nearly 5x cheaper on that same suite. The saving is not a lower unit price: it comes from charging the input once and billing nothing for output, which removes the most expensive part of batch judgement.
The internal example makes the pattern concrete. Gemini Notebook uses it as a real-time intent router, spending a few milliseconds deciding whether a query is routine or complex so that routine traffic takes a light path and only hard cases reach a deep reasoning model. Google reports up to 40% lower latency on everyday queries as a result. That is the template for this class of model: it does not produce the answer, it decides which road the request takes.
Third-party evaluation is thin. Artificial Analysis scores it 66 on the Decision Index v2.1 and does not place it against other vendors, so the number is best read as a position on that vendor's own leaderboard rather than a cross-model ranking.
Ecosystem coverage arrived on day one. OpenRouter served it immediately, and LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit, OpenCode and the google-genai SDK all support it. Support in Gemini on Google Cloud is listed as coming soon through a sign-up form, the only remaining queue.
One thing to settle before use: it returns probabilities, not conclusions. Where the threshold sits, how many options qualify, how the bands are drawn, all of that is your design decision. The model pushes execution cost close to zero and leaves the criteria where they belong, with you.
Key Features
- Three question types in one request: Yes or no returns a probability with optional true and false criteria; choice picks from 2 to 255 options and returns the winner with normalised probabilities for all candidates; score rates against an ordered rubric and returns a weighted score plus per-band probabilities. All three can be submitted together.
- 1,000 questions per request: A single call takes up to 1,000 independent questions, and latency stays nearly flat from 1 to 100, turning per-item judgement into one call and removing the loop and its retries.
- No output charges: $0.10 per million input tokens and zero for output. The shared input is billed once per request, each question only adds its own instruction and criteria, and there are no thinking token, transcription or OCR fees.
- Native multimodal input: Text, images, audio, video and PDFs can share one input up to 256K tokens, so judgements run on the original material without a transcription or recognition pass first.
- Low latency: 150 to 250 ms end to end with median model compute of 50 to 100 ms, measured on contexts up to 16K tokens.
- Calibrated probabilities and confidence: Every answer carries a probability and a confidence score, so you can route by threshold or send low-confidence items to a human instead of forcing a binary call.
- Supported on day one: Live on OpenRouter immediately, with support in LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit, OpenCode and the google-genai SDK.
- Free tier in AI Studio: Prototype on the free tier in Google AI Studio and move to paid traffic later. Gemini on Google Cloud is listed as coming soon through a sign-up form.
Use Cases
- Agent routing, deciding which model handles a request
- Content moderation and compliance, screening hundreds of criteria in one pass
- Eligibility screening for applications and forms
- Candidate ranking and match scoring
- Batch labelling and rating of multimodal assets
- Ticket classification and priority triage
- Automatic grading of eval sets against a rubric
Pros
- Output is unbilled and shared input is charged once, putting batch judgement at $0.10 per million input tokens
- Up to 1,000 questions per call with latency nearly flat from 1 to 100, so per-item judgement becomes one call
- Yes or no, multiple choice and rubric questions can be mixed in a single request
- Takes images, audio, video and PDFs directly, so judgements run on raw material
- Returns calibrated probabilities and confidence scores for threshold routing or human review
- Ships with OpenRouter, LangChain, Vercel and other framework support on day one
Pricing
$0.10 per million input tokens with no charge for output. The shared input is billed once per request regardless of question count, and each question only adds its own instruction and criteria tokens. No output token fees, no thinking token fees, no separate audio transcription or OCR fees. Google AI Studio offers a free tier for prototyping. Support in Gemini on Google Cloud is coming soon and requires a sign-up form.
Summary
Gemini Decision 1 is Google's first decision model, paired with a new Decisions API. It does not write answers token by token: you hand it one multimodal input plus up to 1,000 closed questions and it returns calibrated probabilities for each. Output tokens are not billed, the shared input is charged once per request, and latency stays at 150 to 250 ms. It fits the parts of a workflow that need a verdict but not an explanation, such as routing, screening, ranking and compliance checks. What it will not do is define your criteria: the questions and the rubric are still yours to write.
Version History
- Gemini Decision 1 and the Decisions API launch (2026-11-11): Google released its first decision model, Gemini Decision 1 (gemini-decision-1), with the Decisions API in the Gemini API and AI Studio. One request takes up to 1,000 questions across yes or no, multiple choice (up to 255 options) and rubric scoring, with a 256K token input cap, native support for text, images, audio, video and PDFs, and 150 to 250 ms end-to-end latency. Priced at $0.10 per million input tokens with free output. Google reports it leads a shared benchmark suite covering agent routing, reasoning, ranking and multimodal understanding with the lowest calibration error, and comes out nearly 5x cheaper than Gemini 3.5 Flash-Lite with the Interactions API. Gemini Notebook uses it as a real-time router with up to 40% lower latency on everyday queries. Artificial Analysis scores it 66 on Decision Index v2.1.