Google Gemini Decisions API: Gemini Decision 1 at $0.10/M Tokens Explained

For the last few years, AI has mostly been asked to answer things. Write this email. Summarize this report. Explain this code. That work is real and useful, but it is only the warm-up. The harder question is now on the table: can AI decide?

Google's answer arrived on November 11, 2026, with the launch of the Gemini Decisions API, built around a new model called Gemini Decision 1. The price is the headline for anyone who runs systems at scale: $0.10 per million tokens. That is not a rate aimed at occasional chat. It is a rate aimed at millions of small judgment calls happening quietly inside business software every single day.

This is a bigger shift than another model release. It marks the moment large AI companies stop competing only over who can write the best paragraph and start competing over judgment itself, the thing businesses have always paid people to supply.

What Google Actually Launched

An API, short for application programming interface, is simply a way for one piece of software to talk to another. A Decisions API means developers can plug decision-making directly into systems they already run. A pricing tool. A claims system. A staffing dashboard. A fraud filter.

The point is placement. The model does not sit in a chat window waiting for a human to type. It sits inside the workflow, at the exact spot where money is made or lost. It looks at the situation, weighs the options, picks a path, and explains why.

Gemini Decision 1 is the engine doing that work. And the $0.10 per million tokens figure tells you what Google expects this to be used for. Chat is a burst of a few thousand tokens when someone needs help. Decisions are a constant stream, thousands or millions of calls a day, each one small, each one with a consequence attached. A price that low signals a bet on volume, not on premium conversation.

What a "Decision Model" Actually Is

Most AI today is a prediction machine. Show it a photo and it says "cat." Show it a sentence and it guesses the next word. Show it a customer and it estimates the odds they will cancel.

A decision model does something different, and slightly harder. It takes a messy situation, looks at the possible paths, weighs the trade-offs, and recommends an action. Then it shows its reasoning in plain language.

Think of the gap between a weather app and a pilot. The app tells you the conditions. The pilot uses them to land the plane.

That difference matters because most business value does not live in describing a situation. It lives in acting on one. Knowing that demand is up is data. Deciding how much stock to reorder, and paying the cost if you get it wrong, is a decision.

Why This Market Matters More Than Chatbots

Chat assistants are becoming ordinary. Many companies offer them, they are easy to copy, and their prices keep falling. The real value is moving to the layer where choices carry a bill attached.

Consider where businesses actually lose money. Rarely in badly written emails. Almost always in setting the wrong price, shipping the wrong amount of stock, approving the wrong loan, or staffing a store on the wrong day. Each of those is a decision with a cost.

Whoever supplies the engine that makes those calls, reliably, at scale, with an explanation attached, sits at the center of enterprise software. That is a far stickier position than being the assistant in the corner tab.

It also lines up with what Google already owns. Decision models need three things: enormous amounts of data, heavy computing power, and tight links into cloud infrastructure. Google has all three in-house. The Gemini Decisions API is the company turning those advantages into a product other businesses rent by the token.

Three Ways This Reshapes the Future of AI

1. AI moves from content to consequences

When AI only writes text, a bad output is annoying. You fix it and move on. When AI approves a refund or reprices a product, a bad output costs real money. That changes every design choice behind the system.

The industry will have to shift from "make the answer sound good" to "make the answer be right, and prove it." Accuracy, audit trails, and confidence scores start to matter more than clever phrasing. Evaluation stops being a research exercise and becomes an operating cost.

2. Explanations become a product feature, not a nice extra

No manager will sign off on a system that says "trust me" when a customer's loan is on the line. A decision a business cannot explain is a decision it cannot defend to a regulator, a board, or a customer.

So the winners in this market will be the systems that show their work. Not a full essay, just the factors that mattered, the data behind them, and how sure the model is. That is the difference between a tool people use and a tool people fear.

3. Agents need a decision layer to be useful

AI agents, software that takes multi-step actions on its own, have a well-known weak spot. They can act, but they are not always good at deciding whether to act.

Add a decision layer and the picture changes. Should this support ticket be escalated? Is this refund above the normal pattern? Does this deal need a human before it moves? Those are judgment calls, and a Decisions API gives agents something closer to judgment, wrapped in rules.

In other words, decision models are likely to become the quiet backbone that makes agentic AI safe enough to actually deploy in a business that has something to lose.

What This Could Look Like Inside a Real Business

The uses are not exotic. They are the boring, high-volume calls companies make thousands of times a day.

• Pricing and discounts: deciding what to charge, and when a discount is genuinely worth it.
• Inventory and supply: choosing reorder points when demand is jumpy and shipping is slow.
• Claims and refunds: clearing the straightforward cases and routing the rest to people.
• Credit and risk: weighing who gets approved, with the reasoning on record.
• Staffing: matching schedules to predicted demand instead of last year's calendar.
• Marketing spend: moving budget between channels as results come in.
• Customer support: deciding what to resolve automatically and what needs a human.

Notice the pattern. These are not creative tasks. They are repetitive judgment calls with clear costs and clear feedback. That is exactly where decision models can prove themselves fastest, and exactly why a token price of $0.10 per million changes the math on whether to automate them at all.

The Risks Nobody Should Skip Past

Decision models bring a new kind of risk, and it does not look like the old ones.

A content model that gets things wrong produces one bad paragraph. A decision model that gets things wrong produces a pattern, thousands of small, consistent mistakes that start to look like company policy. A pricing engine that leans slightly too aggressive. A claims filter that shaves a little too fine. An approval model that quietly favors one group over another. These problems do not announce themselves. They hide in the average.

Three guardrails matter most:

• Reversibility. Let the model handle the calls you can undo. Keep humans on the ones you cannot.
• Record keeping. Every decision should leave a trail: what the model saw, what it chose, and how confident it was.
• Regular checks. Models drift as the world changes. What was right last quarter may be wrong this one, and nobody notices unless someone looks.

There is a human question too. When a system makes the call, who owns the outcome? Companies need a clear answer before they deploy, not after a customer complains. "The model decided" is not a defense anyone wants to give in a meeting.

Actionable Steps for Business Leaders

Whether or not you touch the Gemini Decisions API directly, the direction of travel is clear. Here is how to get ready.

• List your decisions. Write down the ten highest-volume or highest-cost calls your teams make each week. That list is your roadmap.
• Sort by stakes. Split them into reversible and irreversible. Start with reversible, prove value there, and only then touch anything that cannot be undone.
• Find the feedback loop. A decision you cannot measure is a decision you cannot improve. If you cannot tell whether a call was good, fix that first.
• Build a test set of real cases. Pull a hundred past decisions with known outcomes. Use them to check any tool before it goes live, and to catch drift months later.
• Set rules in advance. Define limits, price floors, approval ceilings, escalation triggers, and build them in as hard stops rather than friendly suggestions.
• Keep a human in the loop where it counts. Automate the routine, escalate the unusual, and make sure someone is watching the edge cases.

The Bigger Picture

The last few years of AI were about generating. The next few will be about deciding. That is a harder problem, a more valuable one, and a more dangerous one.

What makes the Gemini Decisions API notable is not just the model. It is the framing. Google is not selling a smarter assistant. It is selling a piece of infrastructure for judgment, priced low enough to be embedded in the plumbing of ordinary business operations, where the real volume sits.

The companies that win with this technology will not be the ones that hand over the most decisions. They will be the ones that pick the right decisions to hand over, measure what happens next, and keep a firm grip on the ones that cannot be taken back.

AI that answers is a convenience. AI that decides is a colleague, and it will need the same things any colleague needs: clear limits, a track record, and someone willing to take responsibility when things go wrong.