Google Gemini Decision 1: The AI Model That Decides Without Writing a Word

Google has released an AI model that cannot write a sentence, and that is exactly the point. On November 11, 2026, Google DeepMind launched Google Gemini Decision 1, a model built to make decisions rather than generate text. Instead of paragraphs, it scores the options you give it and returns a probability for each, with a confidence score your software can act on.

Most AI models work by predicting one word at a time until they produce a full answer. That is powerful, but slow and expensive when all you need is a yes or no, a choice from a list, or a score on a scale. Google Gemini Decision 1 takes a different route: it looks at your input once, weighs every candidate answer in a single pass, and returns numbers instead of words. There are no output tokens to pay for, because it generates no text at all.

What Google Actually Launched

Google announced the model on November 11, 2026, in a blog post by DeepMind researchers Min Choi and Felix Yu titled “Gemini Decisions API: Fast, calibrated multimodal decisions”. Chief executive Sundar Pichai also introduced it on X the same day. The model, gemini-decision-1, is available through the new Gemini Decisions API inside Google AI Studio.

For now, it runs only on the Gemini Developer API as a v1beta endpoint. Google Cloud support, through Vertex AI, is listed as coming soon, so large companies needing enterprise data terms must wait. It is also listed on Vercel’s AI Gateway and OpenRouter.

It came in a busy week for Google. A day earlier, the company rolled out Gemini 4 Carbon, its most powerful model, and Gemini 4 Hydrogen, a faster, cheaper model for everyday coding. Google Gemini Decision 1 is the third piece, a faster judge rather than a bigger brain.

A Model That Judges Instead of Writing

A normal language model generates its answer word by word, so cost and waiting time grow with the reply’s length. A decision model skips that step. You send one input, such as a message, a photo or a PDF, with your questions, and it scores every candidate answer in one forward pass.

Because nothing is generated, output tokens are always zero, and you are never billed for them. With a normal model, asking twenty questions about one document means paying for twenty answers. With Google Gemini Decision 1, you ask all twenty in one request and the shared input is billed once. One independent test found that asking twenty questions in one request used over eighteen times fewer tokens than asking them one at a time.

Every answer carries a confidence score between zero and one. When confidence is high, the app can act automatically. When it is low, the case goes to a human reviewer or a larger model. Developers get a number they can build rules around, instead of guessing whether the AI is sure.

The Three Question Types

The model supports three question types, mixable in one request. The first is a yes-or-no check, which Google calls “noul”. You describe a condition, such as whether an image contains a safety hazard, and the model returns a probability between 0.01 and 0.99 that it holds. It never reaches zero or one: the model may never claim total certainty.

The second type is multiple choice. You provide two to 255 options, and the model picks the best one while returning every option’s probability. The probabilities add up to one, so developers can see how close the runner-up was.

The third type is a score against a rubric you define: an ordered scale, say from poor to excellent, for which the model returns a weighted score with each level’s probability. This suits grading tasks, such as rating a support reply. Up to 1,000 questions fit in one request, against a single shared input of up to 262,144 tokens. The input can be text, an image, audio, video or a PDF, and the questions stay independent, so adding or removing one does not change the rest.

Close-up of a humanoid robot with glowing blue lights representing Google Gemini Decision 1
A close-up of a modern humanoid robot with glowing blue lights. Photo by Kindel Media on Pexels.

1,000 Questions, One Bill

Pricing is where it makes its case most clearly. On the paid tier, input costs 0.10 dollars per million tokens and output costs nothing. A free tier for prototypes allows 200 requests per minute and 10,000 requests per day.

Google claims the model is nearly five times cheaper than running the same checks through Gemini 3.5 Flash-Lite, its low-cost generative model, on the company’s own benchmark suite, and claims the lowest calibration error among the models compared. Both figures come from Google’s own testing and have not been independently verified, so read them as claims rather than settled facts.

Google reports end-to-end latency of 150 to 250 milliseconds, nearly flat as questions rise from one to one hundred. Independent testing is more modest: one developer in India recorded about 456 milliseconds for a yes-or-no question, roughly twice Google’s figure. Google’s numbers also apply only to contexts up to 16,000 tokens, not the full 262,144-token ceiling.

What Developers Will Actually Use It For

Google names four main uses, all jobs where a full chatbot is overkill: routing, deciding which tool or workflow handles a request; policy checking, testing whether content breaks a rule before publication; model selection, picking the right model tier; and grading retrieved context, judging whether a database document is actually relevant to a question.

Those examples point to wider uses. A support team could classify thousands of tickets in seconds, sending only uncertain ones to a person. A marketplace could check product photos against its rules on upload. A hiring platform could grade applications and flag borderline cases. The model does the cheap, fast judging; people or larger models do the expensive thinking.

The design fits how modern AI apps are built. Many chain several models together: one retrieves information, another writes the answer, and something must check the result before it reaches the user. Google Gemini Decision 1 is built to be that checker, running on every request without blowing the budget.

How It Compares With OpenAI and Perplexity

Google is not alone in betting that decisions deserve their own model. OpenAI has its own Decisions offering, and the most direct rival arrived a day after Google’s launch, when Perplexity released Decider v1.3 on November 12, 2026, with open weights under the Apache 2.0 licence. Perplexity’s model can run on a developer’s own hardware; Google’s runs only on Google’s servers for now.

Independent benchmarks place it mid-pack. The Artificial Analysis Decision Index, updated on November 16, ranked it 10th out of 19 decision models, with an index score of 65.98 and a speed of about 211 milliseconds per task. Solid for a first release, but not the commanding lead Google’s launch post suggests.

Price is where Google has a clearer edge. At 0.10 dollars per million input tokens with no output charge, it undercuts most rivals for high-volume classification. The trade-off is control: until Vertex AI support arrives, businesses needing to run the model inside their own cloud, or under strict data rules, will find Perplexity’s open-weight model more attractive.

Robotic hand reaching toward a glowing digital network
A robotic hand reaching toward a glowing network. Photo by Tara Winstead on Pexels.

The Limits You Should Know

For all its promise, the model has real limits. The biggest is that it can only judge; it cannot explain itself. Ask why a ticket was classified as a refund request and it has nothing to say. Ask it to extract, classify and verify in three steps, and it cannot, because it makes a single pass with no reasoning chain. Anything needing an explanation still requires a generative model.

Beta status brings practical limits too. There is no caching or batch processing yet, and the v1beta endpoint lacks Google Cloud’s enterprise data terms, regions and support. Regulated companies will likely wait for the Vertex AI version before sending it customer data. The confidence scores have not been validated against external calibration data, so developers should test them first.

There is a subtler risk. Because the model returns tidy numbers, it is tempting to treat them as facts. A probability of 0.92 does not mean the model is right 92 times out of 100 in your use case; it means the model is confident in its own scoring. Teams that wire it into automated decisions without checking accuracy on their own examples could automate mistakes at a speed no human can match.

What Happens Next

The next milestone is Google Cloud support. Once the model arrives on Vertex AI, large businesses can use it under enterprise contracts, in their own regions, with the compliance guarantees they require. That is when it should move from experiments into production at banks, retailers and media companies.

Competition will sharpen the product too. Perplexity’s open-weight Decider is an alternative developers can run anywhere, and OpenAI and Microsoft will not stand still. Expect prices to fall, accuracy to improve, and the first independent calibration studies to test the confidence scores. Caching and batch support would cut costs again for heavy users.

For now, Google Gemini Decision 1 is free to try in Google AI Studio, and the documentation is public. Anyone building an app that sorts, filters, routes or grades content can test it in an afternoon. It will not write your emails or explain its reasoning, but if your app needs a thousand small decisions a second, that may be exactly what you want.

Leave a Comment