Anthropic expanded its OneGov agreement with the US General Services Administration on 27 October 2026, giving government buyers a 40% discount on Claude Enterprise, Claude for Government and the Claude API, according to Anthropic's announcement. If you build software for the public sector, expect more agencies to ask for Claude in the next few months. Treat the discount as a procurement fact, not a security one. Every agency use case still needs its own review.
What is OneGov?
OneGov is the GSA programme that negotiates government-wide terms with technology vendors, so agencies do not each negotiate alone. Anthropic says agencies "can buy Claude at the OneGov discount through GSA's Multiple Award Schedule today," with no minimum spend commitment.
What does the new deal cover?
Here is what Anthropic's post says, point by point.
- Products. A 40% discount on Claude Enterprise, Claude for Government and the Claude API.
- Buyers. Federal agencies under the GSA agreement. Anthropic adds that "the discount also extends to state, local, tribal and territorial governments."
- Channels. Anthropic says that "for the first time, the discount applies across every product surface." Direct purchases now, cloud providers later.
- Savings. Anthropic says the discounts "are estimated to save the government $1 billion." That is Anthropic's estimate, not an audited figure.
How much is government already using Claude?
According to Anthropic, about 500 US federal, state and local government organisations now use Claude. It says usage across federal civilian agencies "has grown about 20-fold in a year," and that weekly active users of Claude for Government are up about eightfold since February.
These are the vendor's own numbers, published to support a sales announcement. They are still useful. They tell you the direction. Public bodies are not experimenting at the edges any more. They are buying.
What security assurances come with it?
Anthropic lists several, and they matter. Claude for Government is at FedRAMP High. On Amazon Bedrock, Claude is authorised for FedRAMP High, with some models authorised for DoD IL4 and IL5 workloads in AWS GovCloud. On Google's Vertex AI, Claude is authorised for FedRAMP High. Claude Enterprise is in process for FedRAMP Moderate through the FedRAMP 20x programme. All of this is from the same Anthropic post.
Notice what those authorisations cover. They say the platform meets a control baseline. They say nothing about whether a given agency's use of the model is a good idea, or whether the data going into it should be there at all.
Why is a discount not a security review?
Because price and risk are different questions, answered by different people.
A 40% discount answers the procurement question: can we afford this, and can we buy it without a new tender? It does not answer the questions a security team asks. What data will go into the prompts? Who in the agency can see the outputs? What happens when the model is wrong about a benefit claim, a permit or a case file? Who checks?
The danger with a good deal is speed. The contract vehicle is ready. The price is lower than last year. The vendor is already authorised somewhere. It all feels done. And the one review that matters, the review of this specific use, gets skipped because everything around it looks approved.
I have seen the same pattern in private companies. A tool gets bought centrally, and then every team assumes someone else checked their use case. Nobody did.
What should the review cover for each use case?
- The data. Which classification of data goes into the model, and does the chosen deployment hold that level of authorisation?
- The decision. Does the output inform a person, or does it decide something about a citizen? The second needs far more scrutiny.
- The owner. A named official who is accountable for the use, not the vendor and not the contract.
- The record. Logs of prompts and outputs, kept under the same records rules as the rest of the agency's work.
- The exit. What happens if the agency needs to stop using the model next year? Can it move its prompts, data and workflows elsewhere?
We wrote about this kind of ownership in AI governance as an operating model. The short version: governance is who decides what, with what evidence, before something ships. A procurement deal is one input to that. It is not the decision.
What does this mean for teams that build for government?
Three practical things.
First, expect Claude to come up more often in public-sector work, including at state and local level, where a 40% discount can make a pilot affordable. Design so the model is replaceable, because the next deal may favour someone else.
That matters most for smaller public bodies. A large federal agency usually has a security office used to reviewing cloud services. A county or a small city may have one person doing IT, procurement and data protection at the same time. For them, a discounted contract on a national schedule can look like the review has already happened. It has not. If you build for those clients, the review is part of the work you offer, not an extra.
Second, check which deployment the agency is actually buying. Claude for Government, Claude on Bedrock in GovCloud and Claude Enterprise come with different authorisations. Your architecture has to match the one they bought.
Third, bring the security review into the project plan from the first week. If it starts after the build, it will be rushed, and a rushed review is the one that misses things.
What would I do this week?
If you work with a public body that uses or plans to use Claude, ask one question: which use cases have been reviewed on their own, separate from the contract? If the answer is "the contract covers it," offer to help write the first proper review.
Where Alongside fits
Reviewing a government AI contract use by use, separate from the discount, is the kind of work covered by Alongside's security and compliance engineering.