The Pentagon Says It Quit Claude. Sources Say It Was There Last Week
A Defense Department official told the BBC the military has stopped using Anthropic's products, five weeks after its own deadline, while people who worked with Claude say it ran inside Palantir's Maven during operations against Iran until days ago.
Eight months after it branded Anthropic a supply-chain risk, the Pentagon says it is finally done with Claude. "The Pentagon has ceased the use of Anthropic products," a Defense Department official told the BBC on Monday, October 5. A department spokesperson gave MeriTalk the same line, using the administration's preferred name: "The Department of War has ceased the use of Anthropic products."
Neither statement says when the last instance was switched off. That gap matters, because the people the BBC spoke to describe a very different timeline.
What the BBC's sources say
According to the BBC, several former U.S. defense officials and AI contractors said Claude was still being used by the department as recently as last week, for research, analysis and intelligence gathering, and in military operations against Iran. One source said Pentagon staff kept using Claude tools, including Anthropic's Mythos models, "throughout the whole period of the controversy."
The sources placed Claude inside the Maven Smart System, the Palantir-operated platform the BBC describes as the Pentagon's primary tool for organizing intelligence and other data. Per the BBC, analysts fed satellite imagery and drone footage organized in Maven into Claude and other large language models for tasks "such as identifying potential military targets," and used it to compile one-page briefs for the chain of command.
Two people also told the BBC that OpenAI tools have recently been adopted more widely in some military departments. Anthropic declined to comment. Palantir has not publicly responded to the report.
The deadline that slipped
When Defense Secretary Pete Hegseth announced the designation in late February, he set a six-month phase-out, putting the cutoff in late August. The BBC says it is unclear why removal took longer.
The Pentagon's own officials had suggested it was nearly finished. On September 11, Emil Michael, the Under Secretary of Defense for Research and Engineering, told DefenseScoop, "I'd say about 90% has transitioned." On Maven specifically he was categorical: "All of the Maven Smart Systems or Palantir work has been transitioned months ago." He said the department was "on track to get it done by the end of the month."
If the BBC's sources are right, the Maven claim was wrong by months, and his end-of-September target was missed. If Michael was right, the sources are describing something other than Maven's core workflows. The two accounts cannot both be fully accurate, and the Pentagon's October statement does not reconcile them.
Michael also made clear what replaces Claude. "The number one idea is to get multiple models," he told DefenseScoop, adding that Maven users who had tried newer versions of ChatGPT "have liked it a lot better than Anthropic," and that Grok was "way faster than both of them." The department has signed agreements with Google, xAI and OpenAI.
Why Claude was hard to pull
The difficulty was flagged in March. Reuters reported then, citing two people familiar with the matter, that Maven used multiple prompts and workflows built with Anthropic's Claude Code, and that Palantir would have to replace the model and rebuild parts of its software. Reuters could not determine how long that would take. Maven-related contracts with the Defense Department and other national security agencies could be worth more than $1 billion, according to that report.
That is the practical lesson of the story. A model is not a light bulb. Once prompts, tool calls and analyst workflows are tuned around one system's behavior, swapping it means retesting everything that depends on it. "Once they become integrated it can be painful to remove them," Lauren Kahn, a former defense official now at Georgetown's Center for Security and Emerging Technology, told the BBC.
The irony at the center
The Pentagon's case against Anthropic, as the D.C. Circuit accepted it on September 25, was that Claude's built-in restrictions made it an unreliable tool for military work. The BBC's reporting suggests the department kept relying on that same tool, in live operations, for weeks after its own deadline to stop. Both can be true: an institution can judge a vendor a long-term risk and still find it too embedded to drop on schedule. But it complicates the picture of a model that could not be trusted when it counted.
It also sharpens the question Anthropic's critics and supporters have argued over since February. Anthropic's position, stated by CEO Dario Amodei on February 26, was that it "cannot in good conscience accede" to an "any lawful use" contract term, and that its two red lines were mass domestic surveillance and fully autonomous weapons. Analysts feeding Maven imagery to a model and compiling briefs for commanders is not, on its face, either of those things. The sources' account does not say the company's restrictions were breached. It says its model was in the loop.
Where this leaves Anthropic
Commercially, the Pentagon's statement changes little. The exclusion was already in force, and the D.C. Circuit's ruling lets the department keep Claude out of its systems and contract work. Politically, the timing is awkward for both sides. The BBC notes that President Trump met Amodei twice at the White House the previous week, and quotes Amodei saying "if we work with the president, we can win safely."
The open questions are concrete. When was Claude actually removed from Maven? Was it used in Iran operations after the August deadline, and under what authorization? And did Palantir's rebuild finish, or did the Pentagon simply stop counting? A department that wants contractors to take its supply-chain designations seriously will need to answer the first of those with a date.