Google launched Gemini Decision 1 on November 11, 2026, and it works differently from every chatbot you know. Instead of writing answers word by word, Gemini Decision 1 scores the options you give it and returns probabilities with confidence scores in about 150 to 250 milliseconds. In addition, one request can carry up to 1,000 questions about the same document, image, video, or PDF,. The price is only $0.10 per million input tokens. Therefore, teams that run classification, routing, and grading at scale finally have a purpose-built tool instead of a repurposed chatbot.
What Is Gemini Decision 1?
Gemini Decision 1 is Google’s first decision model, launched November 11, 2026. It answers closed questions instead of generating text: you send one shared input plus up to 1,000 yes/no, multiple-choice, or rubric questions, and it returns a calibrated probability and a confidence score for each one in a single forward pass. Because it never writes free-form text, it skips the slow token-by-token decoding that makes chatbots expensive for simple judgments. Google DeepMind built it for the small decisions that power real software, such as tool selection, policy checks, model routing, and context grading. For example, Google says Gemini Notebook already uses it to route queries to different models and levels of reasoning effort.
A Decision Model, Not a Chat Model
A chat model reads your prompt and writes a reply one token at a time, which costs time and money. In contrast, a decision model receives a fixed set of candidate answers and scores them all at once. As a result, you get numbers instead of sentences: a probability for each option and a confidence score that shows how sure the model is. However, this design has a clear trade-off. The model cannot explain why it picked an option, and it cannot reason through several steps in one request. Therefore, use it to judge, and use a chat model to explain.
How the Gemini Decisions API Works
The Gemini Decisions API is the endpoint that powers Gemini Decision 1, and it follows a simple pattern. You send one multimodal input of up to 262,144 tokens, which can mix text, images, audio, video, and PDFs, together with 1 to 1,000 independent questions about that input. The model then scores every candidate answer in a single forward pass and returns probabilities plus confidence scores for each question. The endpoint is POST /v1beta/decisions, and the google-genai SDK supports it from version 3.2 with client.decisions.create. Meanwhile, integrations are rolling out across OpenRouter, LangChain, Vercel, and LiteLLM. For example, Gemini Omni 1.2 Flash handles video generation, while this model handles the judgments around it. Finally, note that you cannot swap the model ID into a normal generateContent call, because the Decisions API uses its own request shape.
The Three Question Types
The API offers three question types, and each one fits a different job. Firstly, noul asks a yes/no question and returns a probability from 0.01 to 0.99 plus a confidence score. Secondly, choice picks one option from a fixed list of 2 to 255 choices and reports the probability of each option. Thirdly, score rates the input on an ordered rubric of 2 to 255 levels, and Google recommends 2 to 10 levels for the cleanest results. Each question has its own instructions, and the model never sees your question keys, so write the full question in the instructions field. However, every question sees only the shared input, not the other questions. Therefore, dependent judgments must run as separate requests, with your own code passing the first result into the second.
Gemini Decision 1 Pricing and Free Tier
Gemini Decision 1 pricing is unusually simple: $0.10 per million input tokens on the paid tier, and $0 for output tokens because the model generates zero output tokens. In addition, Google offers a no-cost tier in Google AI Studio for prototyping, with limits of 200 requests per minute, 50,000 input tokens per minute, and 10,000 requests per day. Because the shared input is billed once, the savings grow fast as you add questions. For example, VentureBeat worked out that one 5,000-token request across 1,000 records costs about $0.50, while twenty separate calls that each resend the record would cost $8.10. Google also reports shared benchmark costs of $0.88 with this model, $0.68 with Jev, and $4.14 with Gemini 3.5 Flash-Lite. As a result, Google calls it about five times cheaper than Flash-Lite for the same decision work.
Gemini Decision 1 Benchmarks: Fast, Calibrated, Ninth Overall
Gemini Decision 1 benchmarks tell two stories at once, so read them with care. Google reports that the model leads across its own benchmark suite, which spans agent routing, reasoning, ranking, and multimodal understanding, and it claims the lowest calibration error in the comparison. For example, on JevBench it scored 92.2 percent, ahead of GPT-6 Luna at 90.9 percent and Jev 1.14 at 90.0 percent. In addition, its calibration error was 0.0250, roughly half of Jev’s 0.0494. However, the independent Artificial Analysis Decision Index v2.1 tells a cooler story: it scores the model at 65.98 and ranks it ninth of 18 decision models. Therefore, the honest picture is a strong, well-calibrated model that is not the outright leader on independent tests. On speed, Google reports median model compute of 50 to 100 milliseconds and end-to-end latency of 150 to 250 milliseconds.
Gemini Decision 1 vs GPT-6 Luna and Jev
The decision-model market already has three serious options, and Google’s entry sits in the middle on price. OpenAI runs its own Decisions API on GPT-6 Luna, and TypeSafe sells Jev 1.14 at $0.042 per million input tokens with free output. In contrast, Gemini Decision 1 costs $0.10 per million input tokens, more than twice the price of Jev, but it accepts images, audio, video, and PDFs in one request. On JevBench, Google’s model leads with 92.2 percent, ahead of GPT-6 Luna at 90.9 percent and Jev at 90.0 percent. However, Jev wins on Google’s own shared cost test at $0.68 against $0.88.
| Model | Input price per 1M tokens | Output price | JevBench score | Calibration error |
|---|---|---|---|---|
| Gemini Decision 1 | $0.10 | $0 | 92.2% | 0.0250 |
| GPT-6 Luna (OpenAI Decisions API) | Not disclosed | Not disclosed | 90.9% | Not disclosed |
| Jev 1.14 (TypeSafe) | $0.042 | $0 | 90.0% | 0.0494 |
Choose Google’s model when your inputs mix text, images, audio, video, or PDFs, or when your team already builds on Google AI Studio. Pick Jev when text-only classification dominates your volume and every cent matters. Go with GPT-6 Luna when your stack already runs on OpenAI.
Gemini Decision 1 Business Use Cases That Fit Today
Gemini Decision 1 use cases start with the repetitive judgments that eat agent budgets. Firstly, model routing: an agent can ask which tool to call or which model should handle a task, and Claude Fable 5.2 leads the Remote Labor Index, so routing the hardest tasks to the strongest model saves real money. Secondly, support ticket triage: one request can classify a ticket by topic, urgency, and sentiment, then route it to the right queue in about 200 milliseconds. Thirdly, content policy checks: one pass can score uploads against a checklist of 50 rules instead of 50 separate model calls. Fourthly, rubric grading: a team can score call transcripts or vendor replies against a fixed rubric and get a probability-weighted score with confidence.
For example, coding agents built on models like the Flint coding model can use decision calls to pick the right tool first. Document checks fit well too: invoice validation and compliance checklists ask closed questions about one shared document.
How to Start Using Gemini Decision 1
Getting started takes four steps, and the free tier covers all of them. Firstly, grab an API key in Google AI Studio, where the no-cost tier needs no credit card. Secondly, install or upgrade the google-genai SDK to version 3.2 or later, because older versions do not include the Decisions API client. Thirdly, send your first request to the POST /v1beta/decisions endpoint with one input and five to ten questions, then check the probabilities against answers you already know. Fourthly, scale up the question count only after the confidence scores look honest on your own data. For example, Gemini 4 Carbon opened to all paying developers on November 10, so teams that need long agent runs can pair it with Gemini Decision 1 for fast routing. In addition, independent guides explain the request shape step by step. Finally, test with your hardest real inputs first.
Limits to Know Before You Build
Gemini Decision 1 has real limits, and knowing them now saves a rebuild later. Firstly, it runs only on the Gemini Developer API v1beta endpoint today, with Gemini on Google Cloud support still listed as coming soon. Secondly, the Batch API, context caching, and Flex or Priority inference are all unavailable. Thirdly, question instructions are text-only, and conversation history plus tool-call blocks are unsupported. Fourthly, it cannot reason in steps, call tools, or explain its answers, so multi-step workflows need your own code around it. In addition, Google has not disclosed the parameter count, training data, or architecture.
Frequently Asked Questions
What is Gemini Decision 1?
It is Google’s decision model, launched November 11, 2026. It scores up to 1,000 questions about one shared input in a single pass and returns probabilities with confidence scores instead of generated text.
How much does it cost?
The paid tier costs $0.10 per million input tokens, and output tokens cost $0. A free tier in Google AI Studio allows 200 requests per minute, 50,000 input tokens per minute, and 10,000 requests per day.
How is it different from Gemini 4 Carbon?
Gemini 4 Carbon is a full generative model for long agent runs, while Gemini Decision 1 only judges closed questions and writes no text. In short, Carbon does the work and the decision model decides how it should flow.
Can it explain its answers?
No, it cannot explain its answers. It returns only probabilities and confidence scores, so pair it with a chat model or a human reviewer whenever you need a reason behind a decision.
How many questions can it answer at once?
One request can carry 1 to 1,000 questions about the same input, as long as the shared input plus all questions fit inside the 262,144-token context window.
Is it better than GPT-6 Luna?
On Google’s JevBench test, Gemini Decision 1 scored 92.2 percent against GPT-6 Luna at 90.9 percent. However, the independent Artificial Analysis index ranks it ninth of 18.
Where is it available?
It is available now through the Gemini API in Google AI Studio on the v1beta endpoint. Support in Gemini on Google Cloud is coming soon, and integrations with OpenRouter, LangChain, Vercel, and LiteLLM are rolling out.
Should small businesses use it?
Yes, if your team classifies tickets, routes agent tasks, checks documents, or grades work against a rubric. The free tier needs no credit card, so a small business can test real workloads before it spends anything.
Does it work with images and video?
Yes, one request can mix text, images, audio, video, and PDFs up to 262,144 tokens. For example, you can ask whether a video shows damage, pick the best thumbnail from a list, or score a scanned form against a rubric.
Conclusion: Should Your Business Try It?
Gemini Decision 1 is the first Google model built only to judge, and that focus is its strength. It answers up to 1,000 questions in one pass, costs $0.10 per million input tokens with zero output charges, and returns calibrated confidence scores in about 200 milliseconds. In addition, the free tier lets any team test it today with no credit card. However, it cannot explain itself, reason in steps, or handle dependent questions, so treat it as a fast judge inside a bigger system, not a replacement for your chat models.
