The dollar AI deals ended in a meter, and meters need ceilings

The week

In mid-August 2026, federal agencies were six weeks from the end of a year of near-free AI. They had been handed a powerful new tool for a token price and had used it widely. Now the deal was ending, and nobody could yet tell them what the new bill would look like.

The story began in August 2025. The General Services Administration, or GSA, is the federal government’s central buying office: it negotiates deals that every agency can then use, much as a large company’s procurement team sets up preferred-supplier contracts. Under a program called OneGov, GSA announced agreements that gave agencies enterprise versions of generative AI tools at nominal prices for a year. Generative AI here means chat assistants that can draft, summarize and answer questions in everyday language. OpenAI offered ChatGPT Enterprise for $1 per agency. Anthropic offered Claude for $1. Google offered Gemini for Government for $0.47 per agency.

Those prices were never meant to last, and by mid-August 2026 the end date was in plain sight. Reseller listings, the catalogs that companies selling to government publish, showed all three offers expiring on September 30. GSA had not said what would replace them.

Reporting at the time added two important details. Some subscriptions would not renew automatically, so an agency that did nothing could simply lose access. And renewal for at least one product would be priced per user, the familiar “per seat” model most people know from office software. Months earlier, the outgoing federal chief information officer had said publicly that prices would not stay at a dollar, and that renewal pricing would draw on usage data, meaning what agencies had actually consumed during the year.

The scale made this more than a footnote. GSA later put the number of federal employees with access through its AI agreements at roughly 3.5 million. That is a large workforce that had grown used to tools whose real cost had been hidden by a promotional price.

So the question that week was simple to state and hard to answer: when the dollar deals end, how will agencies pay for something that charges by the amount it is used, and who decides how much gets used?

From a fixed fee to a running meter

The dollar price never decided adoption. What decided it was a security authorization, a ruling on which data could go in, and an owner for records and oversight. A security authorization is the formal sign-off that a system is safe enough to hold government information. The data ruling says what employees may and may not paste into the tool. The records owner makes sure that what people create is kept as the law requires. Those three things, not the price tag, were what let an agency switch the tool on.

A year later, price came back in a different form. Every signal in August pointed toward usage-based or per-seat pricing, and usage pricing changes who controls spending.

A flat fee works like a buffet: you pay at the door, and how much anyone eats does not change the bill. Usage pricing works like ordering from a menu, where every dish adds to the check and the total depends on what everyone at the table orders. Under a flat fee, a contracting officer fixes the cost at award. Under a meter, the cost is set by thousands of employees each time they send a prompt, upload a file or run an agent.

A few terms are worth defining here. A prompt is the instruction or question an employee types into the assistant. An agent is an AI tool that carries out a series of steps on its own rather than answering once, such as searching documents, drafting a reply and revising it, and each of those steps consumes capacity. And most AI providers measure that capacity in tokens, which are small chunks of text, roughly a short word or part of a longer word. Both the text going in (a question, an uploaded report) and the text coming out (the answer) are counted. A long document summarized several times by several people can add up to a great many tokens.

That shift raised three decisions that agencies needed to settle before September 30, and each belongs to a different office. The first is how to fund a meter, which belongs to the chief financial officer. The second is who gets to run it, which belongs to the CIO and program leaders. The third is how to keep alternatives, which belongs to acquisition and records staff. Together they cover what changes when a fixed-price tool becomes a consumption service.

Funding a meter

Federal agencies cannot obligate money they do not have. To obligate money means to make a legally binding commitment to spend it, such as signing an order. The Antideficiency Act forbids obligations in excess of available appropriations, and open-ended consumption pricing sits awkwardly with that rule. An appropriation is the amount Congress has given an agency to spend for a purpose. If a bill can grow without limit depending on how much staff use a tool, an agency could find itself committed to more than it has, which the law does not allow.

Private-sector leaders will recognize a softer version of the same problem. A company can overspend a budget and sort it out later. A federal agency does not have that option, which is why the controls need to be designed in from the start.

Cloud computing has dealt with the same problem for a decade, and the answers carry over. Cloud services, where an organization rents computing power from a provider instead of owning servers, have long been billed by consumption. Agencies learned that an order needs a funded ceiling, a not-to-exceed amount, and a way to see spending against it in close to real time.

In practice that means three controls.

1. A funded cap. The order should cap obligations at a funded amount, with modifications to add funds as use grows. The agency then knows its maximum exposure at any point.
2. Regular visibility. The vendor’s administrative console, or a billing feed, should report consumption by bureau or office at least weekly. The console is the control panel the provider gives administrators; a billing feed is a regular data file of usage and charges that the agency’s finance systems can read.
3. A named trigger. Someone should own a burn-rate threshold that triggers a decision before the ceiling is reached, such as throttling, re-prioritizing or adding funds. Burn rate is simply how fast money is being spent. Throttling means slowing or limiting usage, for example by lowering allowances for some users.

Agencies running under a continuing resolution have extra reason to watch the burn, since they may be funding a new cost at the prior year’s rate. A continuing resolution is the stopgap Congress passes when a full budget is not agreed; it generally keeps funding at the previous year’s levels. An agency in that position may be paying a new and growing bill out of a budget sized for a world where the bill was a dollar. Budget and appropriations law staff should review the structure before the first order.

Model choice is also a cost decision under a meter. Most providers offer a family of models, from small, fast and inexpensive ones to large, highly capable ones. Token prices differ widely between a provider’s largest and smallest models, so the same request can cost many times more on one than the other. Tidying the grammar of a two-line email does not need the most powerful model. Setting a default model for routine work, with the more capable models available on request, is one of the cheapest controls an agency has.

Deciding who runs it

A nominal fee rewarded broad enablement. Giving a tool to every employee cost nothing extra, so many agencies did. A meter reverses that logic. Light users cost little, while heavy users, long documents and agentic workflows can cost a great deal. The access model has to change accordingly.

Agencies can sort users into tiers based on what the first year showed.

• General staff might keep access to a basic model with a monthly allowance, enough for everyday drafting and questions.
• Teams with measured use cases, such as drafting correspondence, summarizing case files or coding, get higher limits and more capable models. “Measured” matters: these are uses where the team can show what the tool saves or improves.
• Pilots of agent features run under separate budgets with their own ceilings, so an experiment cannot consume the money meant for everyone else.

Usage data from the first year is the raw material for this. Most enterprise consoles show active users, volume and feature use. Agencies that exported and kept that data could size tiers from evidence instead of guesswork. Those that did not were left estimating, and estimates tend to be either too generous, which wastes money, or too stingy, which frustrates the staff getting the most value.

Chargeback matters too. Chargeback is the practice of assigning a shared cost back to the units that used it. If a department funds the order centrally, bureaus have no reason to economize. Showing each bureau its own consumption, even without billing it, changes behavior. People manage what they can see. It also gives the chief AI officer a basis for reporting value by program rather than in aggregate. Cost per active user and cost per measured use case, reviewed monthly, are the two numbers a budget officer will ask for first.

Keeping alternatives

The three offers did not have to end the same way, and agencies were better placed if they could move between them. Being able to switch is the strongest negotiating position any buyer has. Two kinds of portability decide that.

Technical portability

The first is technical. Workflows built through a model-agnostic layer, such as GSA’s USAi evaluation platform or an agency’s own gateway, can be pointed at another model when prices change. A gateway is a piece of software that sits between the agency’s staff and applications on one side and the AI providers on the other. Employees and systems talk to the gateway; the gateway talks to whichever model the agency has chosen. Change the provider behind the gateway and, ideally, the people in front of it barely notice.

Custom assistants built inside one vendor’s interface usually cannot be moved that way. If a team spent months tuning instructions and loading reference files into a custom assistant on one platform, that work may have to be rebuilt elsewhere. An inventory of what staff built in the first year, and where, shows how locked in the agency is.

Records portability

The second kind is records. Conversation histories, uploaded files and custom configurations may be federal records, depending on content. Federal records law requires agencies to keep documents that show how government business was done, and a chat in which an official drafted a policy memo can fall into that category. An agency needs to know how to export and retain them if it stops paying, or narrows access, at the end of a term. That should be tested with the records officer before the decision, because retention after a subscription ends varies by product and by contract. Discovering that a year of records cannot be retrieved after the subscription lapses is the kind of problem that ends up in an inspector general’s report.

With exports and a gateway in place, the September decision became a real choice. An agency could renew broadly, narrow use to proven cases, move some workloads to another provider, or compete the requirement. Without them, the only realistic option was to accept whatever terms arrived.

Two Septembers at one department

Take a mid-sized department with several bureaus and several thousand staff who had access to one of the dollar-priced assistants, and run its September two ways.

In the version that goes well, the CFO’s office sets a funded ceiling for the first renewal order and asks appropriations counsel to review it. The CIO’s team exports a year of console data and finds that most staff used the assistant lightly, while a few teams, such as the correspondence unit and a software group, used it heavily and could point to measurable time savings. Those teams get a higher tier. Everyone else keeps a basic allowance on a lower-cost default model. A small agent pilot gets its own budget line. Each bureau receives a weekly report of its own consumption, and a named deputy CFO owns the threshold at which spending triggers a review. Meanwhile, new workflows are built through the agency’s gateway, the records officer has confirmed how to export conversations, and the acquisition team knows it could move some work to another provider. When new terms arrive, the agency compares them calmly.

In the version that goes badly, nobody exports the usage data, so nobody knows who relied on the tool. Access lapses on September 30 for some subscriptions that did not renew automatically, and teams discover it when their work stops. Others renew everyone at the most capable model, funded centrally, so no bureau sees its own cost. A few months later the budget office notices that spending is running far faster than planned under a continuing resolution, and the only quick fix is a blunt cut that hits the high-value teams along with everyone else. The records officer, asked late, finds that some histories may not be retrievable.

Neither version requires unusual technical skill. The difference is whether the three decisions were made deliberately, by the right offices, before the deadline.

Finance, technology, acquisition and records on one page

Running a meter well is not complicated, but it does require people in finance, technology, acquisition and records to work from the same page.

• Every consumption-priced order has a funded ceiling and an owner for the burn-rate threshold. Treating the order like a traditional license, and assuming the cost is known at signature, invites an overrun.
• Access is tiered from evidence, with heavier use reserved for measured use cases, instead of carrying over all-staff access at the most capable model because that is what people are used to.
• Routine work runs on a lower-cost default model, rather than every request going to the most expensive model by default.
• Each bureau sees its own consumption. One central bill that nobody below the department level ever looks at controls nothing.
• New workflows go through a model-agnostic gateway, with an inventory of anything built inside a single vendor’s tools, so lock-in is known before it is time to negotiate.
• Exports and retention are tested with the records officer before any term ends, instead of assuming the vendor will keep everything indefinitely after payment stops.

Outside government, the pattern is the same. Many enterprise AI tools are moving toward consumption pricing, and introductory pricing rarely lasts. The disciplines above, a ceiling, visibility, tiers, a cheaper default, chargeback and the ability to leave, apply to a bank or a hospital system as much as to a federal department. The specific legal rule differs; the management problem does not.

What the budget officer will ask first

A leader does not need to understand tokens in detail to steer this well. These questions will surface most of what matters.

• What is the funded ceiling on our AI orders, and who reviewed it against appropriations law or our own budget rules?
• Who owns the burn-rate threshold, and what exactly happens when we hit it?
• Who are our heaviest users, and what measurable work does that usage support?
• Do any of our subscriptions end without automatic renewal, and who would notice if access stopped?
• If the new terms are worse, what are our real alternatives, and how long would switching take?

Two checks sit underneath those answers. One is whether the first year’s usage data was exported and kept, since every tier depends on it. The other is whether exporting conversations, files and configurations has actually been tried with the records officer, rather than assumed to work.

The dollar deals were a way to get tools into people’s hands. What comes after is a way of paying for them, and that is a management decision, not a technical one.

With hindsight

The terms did arrive, in different shapes.

On September 10, GSA announced a 27-month OneGov offer for OpenAI’s models, effective October 1 and running through 2028. It replaced the flat fee with a 50 percent discount on token-based usage, with no platform fee, no minimum order and no spend commitment. Agencies would pay for what they used, at half the usual per-token rate, without an entry charge and without promising to spend any particular amount. It also opened to state, local and tribal governments, widening the pool of public bodies that could buy on the same terms.

The other two providers moved on a different timetable. By the end of September, Anthropic’s and Google’s offers had been extended into late October and mid-November, buying agencies a little more time rather than settling the question. GSA said it would renew agreements where appropriate.

The design confirmed the meter. The deal that did arrive was exactly the consumption model the August signals pointed toward. And the absence of a spend commitment cuts both ways: it gives agencies flexibility, and it leaves budget control entirely with each agency. No contract term will stop an office from overspending; only its own ceilings, reports and thresholds will.

Published pricing does not tell us how much agencies will spend. That will depend on the ceilings, tiers and exits each one set up, and we have not seen consolidated figures. A discount on a meter is only a saving if someone is watching the meter.

In any sector, this episode is a clean illustration of how technology costs tend to evolve. A low introductory price builds the habit, and then pricing shifts to reflect real use. The organizations that come through well are the ones that treated the free year as a chance to gather evidence, keep their options open and decide in advance who would own the bill. The ones that treated it as free are the ones that later had to make rushed cuts.

The staggered end dates also show why portability matters in practice. When providers move at different speeds and on different terms, an agency that can route work through a gateway and export its records can pick and choose. One that cannot is left waiting on whichever deal arrives last.