Pentagon Says Its Claude Phaseout Is Complete, Sources Saw It Last Week
The Pentagon told the BBC it has ceased using Anthropic products, completing a phaseout ordered after the company refused to remove safety guardrails from military models. Sources said Claude, including its Mythos models, was still in use on Palantir's Maven intelligence platform as recently as last week. A judge had already ruled the blacklist was illegal retaliation.
On this page
The statement, and the sources it contradicts
The Pentagon told the BBC it has ceased using Anthropic products, a statement that closes the most consequential standoff yet between a frontier lab and a government customer over safety guardrails. The claim arrives with an unusual footnote: former defense officials and AI contractors told the BBC that Claude, including its Mythos models, was still in use as recently as last week, for research, analysis, intelligence work, and operations connected to the conflict with Iran, and that removal was slow precisely because the integrations ran deep.
At the center of the usage was Maven Smart System, the Palantir-operated platform that organizes the Pentagon's intelligence data, satellite imagery, and drone footage, where Claude had been fed material for tasks including identifying potential military targets. This blog covered the appeals court ruling that upheld the Pentagon's exclusion in July; the new statement is the operational end of that fight, and a judge had already ruled the underlying blacklisting was illegal retaliation.
How a guardrail fight became a blacklist
The sequence matters for anyone modeling this dispute. Anthropic's tools had been in government use since 2024. In February, the Pentagon demanded removal of Claude's safety guardrails for military workflows; Anthropic refused, citing concerns about mass surveillance and autonomous weapons. Defense Secretary Pete Hegseth labeled the company a national security supply-chain risk on 27 February, one day before US and Israel strikes on Iran, and set a phase-out deadline of late August.
Anthropic called the designation unprecedented and unlawful and sued the administration, and it declined to comment on the Pentagon's new statement. The gap between the two positions is the whole story: the government framed a company's usage restrictions as a supply-chain vulnerability, while the company framed the same restrictions as the reason its models could be trusted anywhere else.
What replaced Claude, and what warms next
The Pentagon has contracted with Google, xAI, and OpenAI in Anthropic's place, which converts the dispute from an abstract policy fight into a live test of whether replacement models accept terms Anthropic would not. Meanwhile the politics have bent back toward warmth: President Trump met Anthropic's chief executive Dario Amodei twice at the White House recently and praised him publicly, even as the company's lawsuit proceeds and the appeals process continues.
The unresolved question is what the government learned about guardrails by switching providers. Anthropic's refusal was specifically about surveillance and autonomous-weapons use cases; whatever commitments Google, xAI, or OpenAI made around the same capabilities are now the operative constraint, and none of them are public. For organizations watching this as procurement precedent, the durable lesson is that model choice now carries geopolitical and legal exposure in both directions, and the guardrails a lab ships are no longer separate from the deal it can win.