Appeals court lets Pentagon keep blocking Anthropic

What the court decided

A federal appeals court in Washington, DC, found that the Pentagon had sufficient grounds to view further deployment of Claude in its information systems as a national-security risk covered by law.

The judges also noted that Anthropic acknowledges building restrictions into Claude that prevent the model from performing tasks its developers intentionally prohibited.

Anthropic spokesperson Danielle Cohen said the company remains confident in its position and is considering its next steps, including asking the full DC Circuit or the US Supreme Court to review the case.

The ruling rejected Anthropic’s claims that the government violated due process and its freedom of speech. The majority treated the dispute as a conventional contract negotiation: the Pentagon excluded Anthropic from its supply chain because the company refused to accept a contract condition the department considered necessary.

The judges said the decision was not retaliation for Anthropic’s support of stricter AI regulation.

Two cases, two different outcomes

Earlier this year, the Pentagon used two separate supply-chain risk laws against Anthropic. The measures were intended to remove Claude from the US armed forces and other federal government bodies by this month.

The legal challenges went to different courts.

The Washington decision was not entirely unexpected. In April, the same panel declined to temporarily block the designation, finding that Anthropic had not met the “strict requirements” needed for immediate relief.

The three judges appeared divided during the hearing before the final ruling. Two ultimately supported the Pentagon’s position; one dissented.

The policy dispute behind the case

Anthropic’s executives had said the company would not allow the government to use its existing AI models to support autonomous weapons or domestic surveillance. US Defense Secretary Pete Hegseth called that position a significant national-security risk.

That framing turned Anthropic’s product restrictions into the center of a government procurement fight. Anthropic argued that the Pentagon had exceeded the authority granted by the supply-chain risk law. The majority instead accepted the department’s view that the refusal to accept a contract condition was enough.

My read is that this is less a judgment on whether Claude is dangerous than a judgment on who gets to set the terms of military access. The court gave the Pentagon room to treat Anthropic’s limits as incompatible with its requirements, even though those limits were part of the company’s stated safety position.

What the ruling does not settle is how the Pentagon is replacing Claude. The department has not published detailed information about its progress with alternative models from:

Employees at Google and OpenAI have opposed their employers’ agreements with the US military, citing ethical concerns and pointing to Anthropic’s earlier refusal to pursue similar cooperation. Neither company changed course, saying support for the US government was important.

Anthropic’s financial exposure

Anthropic reported lost revenue shortly after the restrictions took effect, as customers became wary of working with a company the government had effectively made an outcast. It has not released new data on the financial effect since then.

At the same time, the company has generally reported rising sales in recent months and is moving toward a possible initial public offering at the end of this year.

That timing makes the legal uncertainty more than a procurement problem. The Pentagon and a significant part of the Trump administration can continue avoiding Claude just as Anthropic may be preparing to present itself to public-market investors.

I think the quiet part of the announcement is the absence of any account of what the designation has cost the government in capability, procurement, or model performance. The ruling establishes that the Pentagon can keep Anthropic out; it says much less about whether the resulting alternatives are better suited to the work.