Anthropic announced on 27 October 2026 that it is expanding its agreement with the General Services Administration (GSA), the agency that centralises purchasing for the US federal government. The new deal applies a 40% discount across Claude Enterprise, Claude for Government and the Claude API, with no minimum spend commitment and payment based on consumption. The company puts the estimated saving to the public purse at one billion dollars.
What was announced
The novelty compared with earlier deals is the reach. According to Anthropic's announcement, this is the first time the reduction covers every route to purchase at once: buying directly from the company, available now, and buying through cloud providers, which is due in the coming months.
Nor does the discount stop at federal level. It extends to state, local, tribal and territorial administrations — abbreviated SLTT in American jargon — a far wider pool of buyers with considerably tighter budgets.
The specific purchasing route is the GSA's Multiple Award Schedule, the catalogue through which public bodies acquire goods and services without running a tender from scratch. Anthropic frames it as part of what it calls the OneGov discount.
Alongside the pricing comes a support programme: Anthropic says it will launch in the coming weeks a dedicated service to help agencies deploy Claude and get the most out of it, carrying over the practices it already uses with private-sector customers.
The adoption figures the company gives: close to 500 US public organisations — federal, state and local — now use Claude. Weekly active users of Claude for Government have grown eightfold since February, and use across federal civilian agencies has grown twentyfold in a year.
The cases Anthropic puts forward
The announcement backs the price cut with a handful of examples, all of them recounted by the company itself:
- Utah. Thirty-four developers across ten state bodies were given Claude Code. Eight in ten had a working environment inside half an hour and, by the close of the four-week pilot, 71% said they had saved twenty hours or more.
- Ransomware response. A state's chief information officer says his team contained an attack in eight minutes, against the two or three days it usually took them.
- A flaw the human pentest missed. A federal security team used Claude Mythos and found and fixed, in roughly two days, a vulnerability that would have allowed a critical production system to be taken over. A penetration test carried out by people the month before had walked past it.
- API migration. A state IT developer shifted close to nine hundred API methods onto a different framework in a couple of days. The original plan had allowed a month for the job.
- Documenting legacy. A state IT director ran Claude Code over twelve old repositories to document them; out of it came a map covering 467 database tables and 27 flagged risks.
- SLTT defence. More than half the states, several of the largest cities and some of the biggest critical infrastructure operators went through Anthropic's cyber defence programme for non-federal government.
The uses the company highlights as most common run along those lines: case backlogs, handling records and forms, and modernising old code.
Certifications: what is cleared and what is not
The detail that matters for anyone working with the US public sector sits in the accreditations. Claude for Government — which Anthropic says offers coding and agentic capabilities close to parity with its commercial versions — has been generally available since 30 September with FedRAMP High authorisation, the level required for sensitive unclassified work.
For those who prefer to build their own tooling, Claude is available on Amazon Bedrock and on Google Cloud's Vertex AI. Both carry FedRAMP High; on Bedrock, selected models are additionally authorised for Department of Defense IL4 and IL5 workloads inside AWS GovCloud.
Claude Enterprise, by contrast, is not yet certified: FedRAMP Moderate authorisation is in progress via the FedRAMP 20x programme.
Context
This is not Anthropic's first move towards the US administration. This blog has already covered the 150 million dollars the company put into the Genesis Mission, with Claude entering fifteen federal agencies, and the Cyber Mission, its cybersecurity programme for critical infrastructure with eleven partners. The GSA agreement is the commercial piece that was missing: not a specific project, but the price at which any body can buy.
The argument Anthropic makes is the familiar one in public procurement: tight budgets and long buying cycles. Hence the insistence that there is no minimum spend commitment and that you pay for what you use. In practice, it removes the most awkward objection — committing public money up front for something that may not work out.
What it means
For anyone building or maintaining software in the US public environment, or selling into it, three concrete things change.
First, the discount reaches the API too. This is not just a cut on corporate chat licences: it affects the cost of building your own applications on top of the model, which is where spending piles up once a project grows.
Second, pay-as-you-go with no minimum turns a pilot into something that fits inside a small budget line. The Utah pattern — thirty-four people, four weeks, measuring hours saved — is exactly the kind of trial this pricing structure lets you propose without a fight with procurement.
Third, the certification map marks where you can go today. Sensitive unclassified work, yes: FedRAMP High is available by three different routes. Claude Enterprise over official data, not yet, until the FedRAMP Moderate process closes.
What the announcement does not say
The one-billion-dollar saving is an Anthropic estimate, and the company does not explain what projected spend it is calculated against. Without that baseline there is no way to check it.
All the use cases are the company's own accounts, mostly anonymous: a state IT director, a cybersecurity professional at a large agency. No public reports back them up, and no methodology is given for the Utah pilot — what counted as "hours saved" and who measured it.
The eightfold and twentyfold growth figures are relative, with no starting number. Multiplying a small base and a large base by twenty are two very different stories.
Nor is there any precision on when the discount will be available through cloud providers — "in the coming months" — nor a date for Claude Enterprise's pending authorisation, nor how long the GSA agreement runs or under what terms it can be revised. As for the support programme, all that is known is that it arrives "in the coming weeks": no scope, and no word on whether it costs extra.