GSA's 40% Claude discount is open to state and local governments. Name who watches the meter before you sign
A county that buys a software license knows its bill in January. A county that buys AI by usage finds out in February, after a busy week at the permit counter. The new federal deal makes the second kind of purchase much easier, and much easier to be surprised by.

On October 27, 2026, GSA announced an expanded OneGov agreement giving federal, state, local and tribal governments a 40% discount on token-based Claude usage, with no minimum orders. Consumption pricing is a different kind of purchase from a seat license: the bill follows behavior. Before a council or board approves any AI service priced this way, name a meter owner, set a monthly cap with an alert, and decide what the service does when the cap is hit.
What GSA announced
On October 27, the U.S. General Services Administration said it had expanded its OneGov agreement with Anthropic. The release describes a 40% discount on token-based usage across the Claude models available to the government, including FedRAMP-authorized environments for sensitive unclassified work. There's no minimum order, so agencies pay only for what they use. And the offer isn't limited to federal buyers: GSA says it covers state, local and tribal governments too, whether they buy directly, through resellers or on the major cloud platforms.
That's good news for a small government with a pilot in mind. A county that couldn't justify a license can try something for a few hundred residents without a contract that assumes thousands. GSA also says it will renew agreements where appropriate as offers expire, which means the terms are not permanent.
Why this is a different kind of purchase
Most software a government buys is priced by the seat or by the year. The amount is known, someone approves it once, and finance notices at renewal. Consumption pricing works differently. The bill follows what people and programs actually do.
That's why "40% off" is less informative than it sounds. The release doesn't publish token prices, and a percentage off an unknown rate isn't a number anyone can budget with. Even with the price known, the other half of the equation is volume: how many questions, how long the documents are, how many times a batch job runs.
A discount on a meter you can't read is still a surprise, just a smaller one.
Decision guide
A license and a meter, side by side
| Consideration | Seat license | Consumption meter |
|---|---|---|
| The budget line | A fixed amount you can approve once | An estimate that depends on how often it's used |
| A busy week | Costs the same as a quiet one | Costs more, in proportion to the work done |
| Who notices | Finance, at renewal | Whoever reads the usage report, if anyone does |
| The risk | Paying for seats nobody uses | A bill nobody planned for |
A discount on a meter you can't read is still a surprise, just a smaller one.
Three places usage leaks
Think about where AI use grows without anyone deciding it should.
A staff tool. Fifty employees get access to a drafting assistant. A few use it constantly, most rarely. Usage is lumpy and it grows as people find new uses, which is usually what you want. But the growth is no one's decision.
A batch job. Someone writes a script that reads every permit application from last year and summarizes it. It's useful and it runs once, until it runs again with a bug, or on a larger archive. One careless loop costs more than a month of chat.
A public assistant. A resident-facing chatbot's usage is set by residents. A storm, a deadline or a news story can multiply the questions in a day. An assistant like the one our note on Dakota BOT's early numbers described is the kind of service where traffic isn't yours to control.
None of these is a reason to avoid consumption pricing. They're reasons to put someone's name next to the meter.
What to put in the council packet
If a board or council is being asked to approve an AI service priced this way, a few lines in the packet will save a difficult meeting later.
A meter owner. One person who reads the usage report every month and answers for it. Not "IT". A name.
A monthly cap and an alert. The cap is the most the service may cost in a month, set with the budget office. The alert is the point, well below it, where somebody gets an email. Ask the vendor or reseller how these are set up, since the OneGov release doesn't describe them.
What happens at the cap. For a staff tool, turning off may be fine. For a public assistant, it needs a plain message and a path to a person, such as a phone number or a form. Decide that before the cap is reached, not on the day.
Usage by program. If everything flows through one account, the total tells you almost nothing. Ask for the usage to be tagged by program or department, so the permit office and the library each see their own.
A review date. Six months in, look at what the usage really was and reset the cap. The first estimate is a guess.
Written out, the whole thing can fit in one paragraph of the staff report: "The [department] will run [service] for [six months] under the GSA OneGov agreement. [Name, title] owns usage and reports it monthly to [committee]. Spending is capped at [amount] a month, with an alert to [name] at [half the cap]. If the cap is reached, [what residents or staff see]. Usage is reported separately for [programs]. The service returns to this body on [date] with actual usage and a recommendation." A board member who reads that knows who to call, what the worst month costs and when they'll hear back.
What the discount doesn't settle
The release leaves several things open, so ask rather than assume. It doesn't publish token prices or an end date. It doesn't say how a state's own procurement rules treat a federal agreement, and that's a question for your procurement office or counsel, since rules differ by state and by purchase size. And three ordering routes (direct, reseller, cloud platform) may not feel the same to a small team that has to manage accounts and security settings.
Cheaper access also doesn't answer the design questions, such as what the assistant is allowed to see, where its answers come from, and how a resident gets to a human. Those decide whether the service helps people or just answers quickly.
The practical way to learn all of this is a small, bounded pilot. Pick one program with a clear volume you already know, such as the questions a permit desk answers by email each week. Run the assistant on that one job for a set period, read the usage every week, and compare the bill with the work it replaced or sped up. By the end you have a real rate per request, a sense of how usage moves on busy days, and evidence the council can weigh. That is a better basis for the next decision than any discount percentage.
How the work moves
The meter owner's month
- Week 1
Read usage by program, not just in total.
- Week 2
Compare it with the cap and the forecast.
- Week 3
Ask each program owner what changed.
- Week 4
Report to the budget office; reset the alert.
- Each quarter
Review the cap and what turns off at it.
Where we fit
We help public bodies scope services like this before they buy: starting from one request type that generates repeated calls, deciding what the assistant may and may not do, and building the cost controls into the design instead of bolting them on afterward. That's part of our government and public sector work. If your board is about to approve a usage-priced pilot, talk to us first, and bring the packet.
Related work: Sutton|Place
We build alongside the people doing the work.
Bring us the problem you’re thinking about.