Gemini Enterprise Pricing: Cost per Seat, Google Agentspace Editions and the Agent Token Bill Nobody Models

Most Gemini Enterprise budgets are built the same wrong way. Somebody takes the seat price, multiplies by headcount, adds a contingency, and walks that number into the planning meeting. Then the first real invoice turns up with a subscription line and a separate overage line on it, and the overage line is the one nobody sized, because it is not driven by how many people you employ. It is driven by how hard your agents work, which is the thing the whole project exists to increase.

This is a straight breakdown of what Gemini Enterprise costs in 2026: the seat prices Google actually publishes, what the pooled quota covers before you start paying extra, how the consumption meter behaves, and the third cost that sits outside the Google invoice entirely. Every figure here was read off Google own pages on 16 September 2026, not reconstructed from other pricing articles.

The short answer

Google publishes exactly two seat prices for Gemini Enterprise and one zero. Business is starting at 21 USD per seat per month with 25 GiB of pooled storage and indexing per seat and a 300 seat ceiling. Standard and Plus share a single card at starting at 30 USD per seat per month, with 30 GiB and 75 GiB of pooled quota respectively. Pay-as-you-go carries a 0 USD seat fee for organizations with 20 or more seats and bills tokens, memory, compute and storage at standard rates instead. A Frontline edition exists for deskless staff with 2 GiB and a 150 seat minimum, and Google does not publish its price. Every seat edition bills consumption on top once pooled quota is exhausted and overages are switched on.

What Google publishes, edition by edition

Worth naming the thing Google does here, because it changes how you should read the numbers: Standard and Plus sit behind one starting-at price. Two editions with visibly different quotas, one shared price card. That means 30 dollars is the floor for Standard and Plus is a sales conversation, not a number you can put in a spreadsheet. Business is the only edition with a price that behaves like a price, and even it says starting at.

Business and Plus get priority access to the latest Gemini models, which is a slightly odd pairing until you read it as the entry tier and the top tier both being sold on model freshness. Frontline is the most restricted: it cannot create custom skills and is limited to pre-provisioned agents. Pay-as-you-go requires an invoiced Cloud Billing account.

Google also bundles more into the subscription than it used to. Google Antigravity and Android Studio AI use are now included in a Gemini Enterprise subscription, currently for a limited group of customers and rolling out more broadly. If you were separately budgeting AI coding tools, check whether that line has quietly moved.

Pooled quota is the number that decides your bill

The seat price is easy. The quota is where forecasts go wrong, and it goes wrong because of one word: pooled.

Your storage and data indexing allowance is calculated per seat but shared across the whole project rather than reserved per person. Buy 200 Business seats and you have 200 times 25 GiB in one shared pot. That sounds generous until you notice what fills it. It is not the number of people who log in. It is how much of your company you connect, because the value of the platform is proportional to how many systems it can see, and every connected system has to be indexed before an agent can search it. A single well-stocked document repository, a ticketing system with ten years of history and a couple of shared drives will eat a pot sized for hundreds of seats without a single unusual user doing anything unusual.

That is also why the quota math tends to arrive late. Connector work runs on the platform team timeline, so the indexing bill lands a quarter after the seat bill, once someone has finished the unglamorous work of getting your apps, APIs and databases talking to each other in the first place. Budget the connectors as a project, and size the pooled quota off the corpus you intend to index rather than off headcount.

When the pool is empty, one of two things happens. Either feature usage stops, or, if an administrator has enabled overages, usage continues at pay-as-you-go rates and the meter starts running. Overages are supported on Standard, Plus and Standard Emerging Market. They are the line item that turns a predictable subscription into a variable cost.

The consumption meter, and why it has no natural ceiling

Here is the structural point, and it is the one to take into the business case. A seat price is bounded by headcount. You cannot accidentally hire four hundred people. Token, memory and compute consumption is bounded by agent activity, and agent activity is the metric every AI program is trying to grow. The success of the rollout and the size of the bill are the same number.

Pay-as-you-go makes this explicit by removing the seat floor entirely. Zero dollars per seat, pay for what you use. For an organization with spiky or uncertain usage that is genuinely the honest pricing model, and Google deserves credit for offering it. It also means the only thing standing between you and an unbounded invoice is a ceiling you set yourself.

Google shipped two ways to soften this on 26 August 2026. Flexible Savings Plans let you commit to a monthly spend and take 10 percent off token costs on a one-year term or 20 percent on three years, with no minimum and no maximum. Note what that does and does not do: it lowers the unit price, it does not put a lid on the total. A cheaper meter running without a ceiling is still a meter running without a ceiling. Google also previewed deferred execution pricing, where eligible agent workloads are scheduled into off-peak windows at up to half the inference cost. That one is worth planning for, because a lot of agent work (overnight reconciliation, batch enrichment, scheduled research) has no reason to run at peak.

How to actually set the ceiling

This is the part that most pricing write-ups skip, and it is the most useful thing in the whole exercise. Since August 2026 you can set a hard monthly cap, and it genuinely enforces.

An administrator opens Gemini Enterprise, goes to Usage and Spending, opens the Usage tab, and works through two separate settings. Under Feature usage there is an Overage toggle with a checkbox per edition, which decides whether users can exceed pooled quota at all. Under Project monthly spend limit there is a Set limit button, which hands you to Cloud Billing to configure a budget with Vertex AI (aiplatform.googleapis.com) selected from Services. That budget covers the Gemini Enterprise app, the Gemini Enterprise Agent Platform and AI coding tools such as Antigravity. Both settings require an invoiced Cloud Billing account and at least one active, non-trial subscription.

Three things about that cap belong in your forecast. It is a hard stop: Google states that when the project reaches the limit, overage usage is automatically stopped and agent API calls pause, with users seeing a Usage limit reached message. It is not exact: Google warns that because stopping usage can take a few minutes to take effect, you might incur charges that exceed your limit. And it is project-scoped, which is the one that bites. Email alerts arrive at 50, 80 and 100 percent of the budget.

What project-scoped really costs you

The cap is set on a Cloud Billing budget, and a Cloud Billing budget can be filtered to projects, services, labels, resource ancestors, subaccounts, credit types and a time period. That is the complete list. There is no dimension for a user and none for an agent.

So the cap is all or nothing across the project. If you run a dozen agents in one project and one of them gets stuck in a retry loop at two in the morning, the ceiling fires for all twelve. Your invoice processing agent and your support triage agent stop because a research agent misbehaved, and they stop wherever they happened to be in whatever they were doing. There is no way to give the experimental agent a small allowance and the production one a large one, and no way to ask the bill afterwards which agent was responsible.

There is a workaround, and it is worth doing even though it is blunt: separate projects for separate risk profiles. Put exploratory or high-variance agents in their own project with their own modest cap, and keep the production agents somewhere their neighbors cannot stop them. It costs you some administrative overhead and it is the only per-agent isolation the billing model supports. We walk through the full measurement of what the cap can and cannot scope to on our Gemini Enterprise spend controls page.

The third cost, and it is not on the Google invoice

Seats and consumption are the two lines finance expects. There is a third and it is usually the largest, which is why it is worth naming before anyone signs.

The point of an enterprise agent is that it does things. Once an agent holds a tool that reaches a supplier portal, a payables system, an ads platform, a SaaS billing API or a checkout, it can move real money. A renewal it decides to approve, an ad budget it tops up, a batch of supplier payments it releases, a metered API it calls ten thousand times: every one of those is a real dollar amount, and none of it appears on your Google bill. Google charges you a fraction of a cent for the inference that made the decision. The purchase itself arrives on somebody else invoice, next month, with no connection to the agent that caused it.

The project spend limit cannot help here, and no amount of tuning will make it. It is filtered to a Google Cloud service, so it only ever sees money that flows through Google Cloud billing. Spend that never enters that system is invisible to it by construction.

The practical consequence for a business case is simple. Forecast three lines, not two. Seats are fixed and easy. Consumption is variable, capped by the project limit, and the cap is worth setting on day one rather than after the first surprise. Agent-caused external spend is variable, uncapped by anything Google ships, and needs a control that lives between your agents and your money: a budget that aggregates per agent across a month, a counterparty rule so an agent allowed to renew one vendor cannot pay a different one, and an approval threshold that holds anything large for a named human. That is what agent spend controls are for, and it is a separate layer from anything in the Gemini Enterprise price list.

A sanity check before you sign

Four questions, in the order they usually turn out to matter.

What are you indexing, in gigabytes? Size the pooled quota against the corpus, not the headcount. Get this wrong and overages start in month two.

Seats or pay-as-you-go? If usage is steady and broad, seats are cheaper and easier to defend. If it is spiky, concentrated in a few teams, or genuinely unknown, the 0 dollar seat fee is honest pricing and you should take it, with a cap set the same day.

Have you set the project monthly spend limit? It is two settings in Usage and Spending and it is the difference between a bad month and a bad quarter. Set it before the first agent goes live, not after.

Can any of your agents cause money to move outside Google Cloud? If yes, the Google cap does not cover your largest exposure, and you need a policy layer in front of the payment rail. If no, you are fine with what Google ships, and you should stop here rather than buying something you do not need.

Gemini Enterprise is a strong product and it now has the best consumption ceiling of any major agent platform we have measured, ahead of Salesforce Agentforce, which alerts rather than refuses, and ahead of AWS AgentCore, which caps a single session with no cumulative total behind it. The pricing is defensible and the caps work. Just price all three lines, and remember that the one Google cannot see is the one with your bank account behind it.