Anthropic has moved Claude for Government out of public beta and opened it to agencies at both the federal and state level, the company said in a post on its own site dated 30 September. The product has been in testing since July. It runs inside an environment holding FedRAMP High authorization, the top tier of the federal program that vets cloud services for sensitive government data.
Anthropic describes the offer as the same capabilities its commercial customers get, minus the compliance trade-offs. New features, the company says, generally follow the commercial release schedule. That is a notable promise for agencies used to waiting months or years for approved versions of tools that private companies already use. It is also a promise only Anthropic can verify, since the post offers no timeline for how closely the two will stay in step.
On the desktop, Claude works directly with files, and jobs such as drafting memos, reviewing contract proposals, and handling casework can lean on Claude’s skills, plugins, and project features. Claude Code, Anthropic’s coding agent, is pitched at teams that build or modernize the software behind public services. Two further products are rolling out in early access under the same controls: a terminal-based version of Claude Code and Claude for Microsoft 365. Agencies must contact Anthropic’s public sector team to join those.
The pricing model is the most concrete detail. Anthropic says there are no seat fees. Agencies instead purchase usage in set blocks under a ceiling that cannot be raised mid-stream, so spending cannot pass the amount an agency has formally obligated. For a government buyer, that matters because overspending on an open-ended bill can trigger legal trouble, not just a budget argument. Administrators can set limits on spending and models for each group of users, track usage by person and by model, and receive alerts as a balance runs down.
The administrative design follows how large departments are built. Department-level administrators can hand out prepaid usage to sub-agencies, which then manage their own users. Agencies can connect the login system they already run, with single sign-on configured in the admin portal. Through SCIM group mappings, each tier of seat gets its own rate limit, dollar cap, and list of permitted models. Defaults can be layered, including what Claude is allowed to connect to.
On oversight, Anthropic says every administrative action is logged, with the trail open to administrators inside the customer organization, and that sensitive operations on its side need approval from two people. Usage exports contain metering data only, which the company says lets agencies respond to security-authorization and inspector-general requests without moving sensitive material around. Conversation history stays on the agency-managed device. Security teams also get documentation to support the agency’s ATO, the formal sign-off that lets a government system go live.
All of this comes from Anthropic’s own announcement. The post cites no named agency customers, no usage figures from the beta, and no independent audit of the controls it describes. Those gaps are normal for a launch post, but they leave open how many agencies actually used the beta.
Procurement is the quiet part of the story. Anthropic says contracting officers can buy from it directly, and that agencies need no separate cloud-provider relationship to start. Rivals that reach government mainly through a cloud reseller will notice. For agencies already piloting AI tools, the practical question for the next quarter is whether a usage-capped, direct contract with a model maker is easier to approve than the reseller route they use today.
Reported by Anthropic (claude.com) on 30 September 2026.
