Google launches Gemini Decisions API, powered by its new Gemini Decision 1 model, to return predefined answers with confidence scores, at $0.10/1M input tokens (Google)

Google's Gemini Decisions API: The AI That Stops Chatting and Starts Choosing

By AI Worker Now Editorial · Published November 12, 2026 · Updated November 12, 2026

For the last few years, the story of artificial intelligence has been a story about words. Bigger models wrote longer essays, answered more questions, and generated more code. The pitch was always the same: talk to the machine, and the machine talks back.

On November 11, 2026, Google changed the pitch. The company launched the Gemini Decisions API, a developer tool built on a new model called Gemini Decision 1. Instead of returning a page of generated text, the API returns a predefined answer, one of a set of choices the developer supplies, along with a confidence score that tells you how sure the model is about that pick. And it does this at $0.10 per million input tokens.

That is a small announcement with a very large shadow. It signals a shift in what the AI industry thinks you actually want from a model: not a conversation, but a decision.

What Google Actually Shipped

Strip away the branding and the Gemini Decisions API is remarkably simple to describe. You give the model a situation and a set of allowed answers. The model returns one of those answers, plus a number that represents its confidence in that choice.

That is the whole idea. No open-ended paragraph. No creative riffing. No chance that the model invents a category you never defined or answers a question you didn't ask.

Three details matter here, and each one tells us something about where the industry is heading:

Predefined answers. The output space is constrained by the developer, not the model. The model's job is to pick, not to write.
Confidence scores. Every decision arrives with a graded sense of certainty. That turns a black-box guess into something a system can act on differently depending on how sure it is.
A published price. At $0.10 per million input tokens, Google is signalling that decision-making calls are meant to run at enormous volume, not a handful of times a day, but millions of times an hour.

Why "Decisions" Is a Bigger Deal Than It Sounds

Here is the problem that has quietly held back AI adoption inside real companies: generative models are wonderful at producing something, and terrible at promising it will be the right something. A support bot that writes a beautiful reply is useless if it routes a furious enterprise customer to the wrong department.

Businesses do not actually run on prose. They run on branches. Approve or deny. Route to billing or route to tech support. Flag for review or let it through. This ticket is urgent, that one is not. Every one of those is a decision with a small, known set of outcomes, and until now, teams had to bolt that behaviour onto a text generator with prompting tricks, JSON formatting, and post-processing code that could break at any moment.

The Gemini Decisions API moves that structure from the developer's duct tape into the product itself. The model is told, in effect: these are the only answers that exist; choose one. That is a fundamentally different contract between developer and model.

The confidence score is the real headline

Constrained answers are useful. Confidence scores are the thing that makes them safe to deploy.

Think about how a well-run human team handles a hard call. A junior employee makes routine decisions alone. A murky case gets escalated to a manager. A truly unusual situation goes to a committee. The trigger for escalation is never the answer itself, it's the uncertainty around it.

A confidence score gives software the same instinct. A high-confidence answer can flow straight through an automated pipeline. A low-confidence answer can be held, sent to a human, or passed to a bigger and more expensive model for a second look. That single design choice turns AI from a system that must be right every time into a system that knows when it isn't, which is a far more realistic thing to ask of any intelligence, machine or otherwise.

It also creates something businesses have been desperate for: an audit trail. If a model says "deny this claim, 0.94 confidence," a compliance team can log that number, track it over time, and set thresholds. Compare that to a paragraph of generated explanations, where the reasoning is persuasive but unmeasurable.

The Price Is a Strategy, Not a Footnote

At $0.10 per million input tokens, the Gemini Decisions API is priced for a world where AI is called constantly and invisibly, the way a database is queried or a cache is checked. Nobody builds a product that checks a cache ten million times a day unless the per-check cost rounds to nothing.

This is the economics of the "decision layer." A decision layer sits between messy human input and the rest of your software stack: an email arrives, a form is submitted, a sensor fires, a customer types something ambiguous. Something has to read that and turn it into a clean label your systems can act on. Today, that job is done by clumsy rules, regex, and humans squinting at screens. If a model can do it for a fraction of a cent at massive scale, the entire architecture of business software changes.

It's also worth noticing that the announcement centres on input pricing, which fits the design. The heavy lifting here is reading the situation. The output is one label and one number, short by definition. Google is pricing the part of the job that scales with complexity of context, not with length of answer.

What This Means for the Future of AI

Three larger trends converge in this launch, and together they sketch out where the next phase of AI competition will be fought.

1. The end of the "chatbot" as the default interface

The chat box was a useful bridge, it let people interact with models before anyone knew what else to do with them. But a chat box is a terrible interface for software. Software needs structured input and structured output. APIs like this one acknowledge that most AI calls will never involve a human reading a response. They will be machine talking to machine, in a narrow vocabulary, at high speed. The visible chatbot becomes a thin customer-facing shell over a boring, reliable decision engine.

2. Reliability becomes the product

For years, the scoreboard was raw capability: reasoning, coding, long context, benchmark scores. Those still matter, but the buying decision inside a company is increasingly made by people who care about failure rates, not peak performance. Constrained outputs and confidence scores are reliability features. They say: this model will not surprise you, and when it's unsure, it will tell you. Expect every major provider to be judged on this axis next.

3. AI starts to look like infrastructure

Infrastructure is cheap, boring, metered, and everywhere. When a decision call costs roughly nothing per million tokens, AI stops being a project with a budget line and becomes a default assumption in architecture. That's the moment adoption stops being a question of whether and becomes a question of where.

How Businesses Will Actually Use It

Because the output space is predefined, the natural use cases are the ones companies already handle with rules engines, spreadsheets, and tired human beings. A few obvious ones:

Support triage. Sort incoming tickets into a fixed set of categories and priorities, and escalate the low-confidence ones to a person instead of guessing.
Document classification. Route contracts, invoices, claims, and forms into the right workflow with a measurable certainty threshold.
Risk and compliance flags. Decide whether a transaction, message, or listing needs review, with a confidence number that satisfies an auditor.
Workflow branching. Let agents choose the next step from a defined menu of actions rather than improvising one.
Lead and intent scoring. Bucket enquiries into defined segments before a human ever opens the inbox.

Notice what these have in common: the categories already exist, the cost of a mistake is real, and volume is high. That's exactly the profile this API is built for.

The Limits Worth Watching

None of this makes the hard problems disappear, and it's worth being honest about the trade-offs.

Someone still has to define the answers. A constrained model cannot tell you that your categories are wrong. If your menu of options is badly designed, the model will confidently pick a bad option. The quality of the decision is capped by the quality of the choices you offer, which puts a premium on careful product thinking, not just prompt engineering.

Confidence is not truth. A score tells you how sure the model is, not whether it is right. Teams will need to calibrate those numbers against real outcomes before trusting a threshold. A model that says 0.9 and is right 60% of the time is worse than useless if you've wired 0.85 to auto-approve.

Narrowness has a cost. The same constraint that makes decisions safe makes them inflexible. When the world shifts and a genuinely new category appears, a system built on fixed answers will cram it into the nearest available box unless someone notices. Monitoring what falls into your "other" bucket becomes a core operational job.

Actionable Insights: What to Do Now

For teams watching this space, the useful moves are less about chasing a specific API and more about preparing for the pattern it represents.

Audit your decisions. List the places where humans or brittle rules currently sort, route, or approve. Those are your candidates, and there are probably more of them than you think.
Write down your answer sets. For each one, define the allowed outcomes clearly. Ambiguous category lists are where these systems fail.
Build for escalation from day one. Decide what happens to a low-confidence answer before you deploy, not after. A human review queue is a feature, not an admission of defeat.
Log confidence, not just outcomes. Track scores over time and compare them against what actually happened. That data is how you earn the right to raise your automation threshold.
Price the decision, not the model. Do the arithmetic on volume early. Cheap per-call costs change which problems are worth solving at all.

The Bottom Line

The Gemini Decisions API is not the flashiest thing anyone will ship this year. It doesn't write poetry or pass an exam. What it does is far more commercially significant: it takes the messy, probabilistic output of a large model and squeezes it into the shape that software has always needed, a known answer, plus a number that says how much to trust it, at a price that makes calling it a million times a background detail.

That's the tell. The most important AI products of the next few years probably won't be the ones you talk to. They'll be the ones quietly choosing, billions of times a day, behind the interfaces you already use, and telling you honestly when they'd rather someone else decided.

TLDR: Google launched the Gemini Decisions API, powered by its new Gemini Decision 1 model, on November 11, 2026. Instead of writing free-form text, it returns one predefined answer from a set the developer supplies, plus a confidence score, priced at $0.10 per million input tokens. The shift matters because it turns AI from a chatbot into plumbing: constrained outputs reduce surprises, confidence scores enable sensible escalation to humans, and the low price makes high-volume decision-making affordable. Expect the industry's scoreboard to move from raw capability toward reliability, auditability, and cost per decision, and expect businesses to deploy these systems first in triage, classification, and workflow routing, where the categories already exist and mistakes are expensive.