Google's Gemini Decision 1 Turns AI Into a 10-Cent Judge
Google's Gemini Decision 1 scores up to 1,000 choices in a single pass for $0.10 per million tokens. It is the first production decision model from a major lab, and it targets the boring work that eats AI budgets.

A model that answers instead of writing
Google has a new kind of AI model, and it does not write essays, code, or poems. Gemini Decision 1, announced on November 11, takes a list of choices and picks one. It is a decision model, built for the unglamorous work that sits between a user asking for something and a system doing it: classification, routing, moderation, ranking, and triage. Google announced it at its AI Engineer event, with Sundar Pichai and Logan Kilpatrick presenting it as the first production decision model from a major lab.
The idea is simple enough that it is surprising nobody shipped it sooner. Most AI work today uses a general model to generate text, then parses that text back into a decision. That is slow, expensive, and fragile. A decision model skips the text entirely. You hand it up to 1,000 candidate answers in one pass, and it scores them. The output is a choice, plus a confidence score you can actually use to decide what happens next.
What Gemini Decision 1 costs and how fast it is
The pricing is the part that will get developers to try it. Google charges $0.10 per million input tokens, and nothing for output, since there is no output text to speak of. Latency is around 200 milliseconds end to end. For comparison, running the same classification through a flagship chat model costs many times more and takes longer, because the model has to generate its answer token by token before you can read it.
The model is multimodal, so the choices it scores can include images as well as text. It is calibrated, which matters more than it sounds: a calibrated model means a 90 percent confidence score is right about 90 percent of the time. That is what lets you automate the easy cases and send only the uncertain ones to a human or a bigger model. Without calibration, confidence scores are decoration.
Google says the project started as an internal DeepMind prototype and went from idea to production API in about six weeks. That speed is itself a signal. Decision models are not a research moonshot; they are a product built from parts Google already had. Anthropic has talked about similar ideas, but Google got there first with a real API, and the Gemini Decisions API is live now.
Where a decision model actually helps
The honest use cases are boring, and that is the point. A support inbox that routes ten thousand tickets a day. A content filter that has to decide, in milliseconds, whether a post breaks a rule. An agent that has to pick which tool to call next, or which of fifty database rows answers a question. A shopping search that ranks results before a bigger model writes the summary. In all of these, the expensive model was doing clerical work. Gemini Decision 1 takes the clerical work off its plate.
This is also where the AI agent wars are heading. Agents that call tools all day spend most of their budget deciding, not doing. A cheap, fast decision layer makes agents cheaper to run, which is what turns demos into products. Expect every major lab to ship something like this within months. Google just set the price: ten cents per million tokens.
The limits worth remembering
A decision model is only as good as the choices you give it. If the right answer is not in the list, it will confidently pick the least wrong one. That makes the design of the candidate list the real work, and it is work that does not disappear just because the model got cheaper. Calibrated confidence helps, but it is not a guarantee: a model can be well calibrated on average and still wrong on the one case that matters.
There is also a question Google has not answered in detail yet: how the model behaves on decisions where the training data itself is skewed. A moderation model trained on one community's norms will misjudge another's. A hiring filter will inherit whatever the historical data says. Cheap decisions made at scale amplify those mistakes faster than expensive ones did.
Still, the direction is right. The industry spent two years making models bigger and more general. The next step is making them cheaper and more specific. Gemini Decision 1 is the first production example of that shift from a major lab, and its price will force everyone else to respond. If you run anything that classifies, routes, or ranks, this is worth a test run this week. The cost of trying it is, quite literally, ten cents per million tokens.
The bottom line
Google beat its rivals to a production decision model, and it priced it like a utility. Gemini Decision 1 will not write your emails or code your app. It will quietly decide which email goes where, which post gets flagged, and which tool your agent calls next, faster and cheaper than the models doing that job today. That is not a headline feature. It is infrastructure, and infrastructure is where the real money gets saved. The Gemini 4 Carbon launch last week was about raw power. This one is about cost. Both matter, but cost is what decides which AI products survive.


