Claude for Government exits beta with agency spending cap
Anthropic launched Claude and Claude Code for US agencies in FedRAMP High, with no seat charge and direct contracting, impacting integrators’ resale model.

Anthropic officially launched Claude for Government on Wednesday (30), a version of Claude and Claude Code operating in a FedRAMP High environment and available to federal and state agencies in the United States. The product had been in public beta since July. According to the company's announcement, agencies will now receive features equivalent to commercial clients, at the same release pace, without compromising compliance requirements.
The announcement is short on numbers and long on purchasing mechanisms. It is precisely here that it attracts those selling technology to the public sector.
No seats, with strict spending cap
Anthropic has eliminated user license fees. Agencies pay for consumption in fixed increments, with a "not-to-exceed" cap that cannot be surpassed. Administrators define user brackets with spending and model limits by group, monitor usage per person and model, and receive alerts on consumption balances. Department-level administrators distribute pre-paid usage to sub-agencies, and group mappings via SCIM set rate limits, dollar caps, and approved models.
This choice responds to a concrete problem for public buyers: a closed annual budget, with no room for variable billing. The seat model, adopted by Microsoft for Microsoft 365 Copilot at $30 per user per month, creates predictable costs but pays for idle licenses. Pure consumption, the standard for APIs, is efficient but intimidates controllers. The strict cap attempts to combine the predictability of one with the efficiency of the other.
There are also controls designed for auditors. Administrative actions are logged, sensitive operations require approval from two people, usage exports include only measurement data, and conversation histories remain on devices managed by the agency itself. The company claims to provide documentation for the ATO process, the operation authorization that each federal agency must issue before deploying a system into production.
Integrator loses a step
The most sensitive point for consulting firms lies in one sentence: procurement officials can contract directly with Anthropic and award general availability terms, without needing a separate relationship with a cloud provider. Until now, the most common federal path to Claude went through Amazon Bedrock, which obtained authorization for FedRAMP High and DoD IL4/5 loads, or through the OneGov agreement announced by GSA in August 2025, which offered Claude to agencies across all three branches for $1 until August 2026.
With the direct contract, the license resale margin diminishes. What remains for the integrator is the work that Anthropic does not handle: connecting Claude Code to legacy systems that support public services, redesigning casework flows, reviewing RFPs, and preparing each agency's security documentation. Claude Code in command line and Claude for Microsoft 365 enter early access in the same environment, with the same administrative controls, which expands the scope for software modernization projects, not distribution.
The same debate, in another currency
Outside the U.S., the push for AI in the public sector is framed in the language of sovereignty, not compliance. In Germany, Cohere signed a definitive merger agreement with Aleph Alpha on September 16, in a deal valued at about $20 billion according to the companies, with dual headquarters in Toronto and Berlin. The Schwarz Group, owner of Lidl, committed $600 million in the next round for Cohere, and the combined company aims to deliver sovereign AI over STACKIT, the group’s cloud.
Both movements target the same buyer from opposite sides. Anthropic believes that a public agency will accept an American supplier if they receive auditing, spending caps, and local data. Cohere bets that European governments will demand that the provider and infrastructure be from the continent. For a consulting CIO with public accounts on both sides of the Atlantic, this means maintaining two competency stacks: integration with FedRAMP environments in the U.S. and with sovereign clouds in Europe, each with its own certification cycle.
Anthropic has not disclosed how many agencies used the beta nor pricing per increment. The absence is significant: without knowing the size of the minimum increment, it is impossible to compare the effective cost with that of a seat contract. However, the most useful information for the services market is already in the text: the manufacturer has decided to sell independently, and the integrator that relied on resale will need to prove its value in the code it delivers, not in the invoice it passes on.