Pentagon to finish cutting Anthropic loose by month’s end

Emil Michael, Under Secretary of Defense for Research and Engineering, told a small group of reporters on September 10 that the Pentagon has moved about 90% of the military’s classified AI workloads off Anthropic’s models. He said the rest should be done by the end of the month. Several frontier vendors are replacing Claude, and Michael said relying on any single model was a mistake whether or not there had been a dispute with Anthropic.

“I’d say about 90% has transitioned,” Michael said, according to DefenseScoop. “All of the Maven Smart Systems or Palantir work has been transitioned months ago, and we’re on track to get it done by the end of the month.” He spoke onstage and on the sidelines of the National Defense Industrial Association’s Emerging Technologies for Defense Conference.

The contract that broke

Claude was hard to remove because it got into classified networks first. A July 2025 contract made it the first frontier model approved for use on classified government networks, and by the time the dispute began its capabilities were deeply built into military systems and those of other major vendors. According to Mayer Brown, the department later asked Anthropic to waive two restrictions in that contract: one on mass domestic surveillance and one on fully autonomous weapons systems. Anthropic refused. The Pentagon’s position was that defense software must be available for “all lawful purposes” without constraints.

A timeline from Tech Policy Press shows how fast it ended. On February 24, Hegseth gave Dario Amodei a deadline of 5:01 p.m. on Friday, February 27, to allow unrestricted use. On February 26, Anthropic said it would not budge. The next day Trump directed federal agencies to stop using Anthropic’s products, and Hegseth designated the company a supply chain risk. Mayer Brown says the directive set a six-month phase-out period for federal agencies, and that letters dated March 3 applied the designation to all Anthropic affiliates, products and services. The firm calls it the first such designation ever applied to an American company.

One model was “the mistake”

“One of the mistakes, independent of whether we had a dispute with Anthropic, is the fact that the Pentagon was solely reliant on one model,” Michael said at the roundtable. “I mean, what if that model were Gemini or some other model that hadn’t kept up with the frontier, right? We would be putting our warfighters in a disposition that was suboptimal.”

His fix is to run several models, because “every month this industry changes — maybe every week — who’s leading and what.” He said the move away from Anthropic has “been great.” Warfighters on Palantir’s Maven who had seen newer versions of OpenAI’s ChatGPT “liked it a lot better than Anthropic,” he said, and they found xAI’s Grok “way faster than both of them. So they each have different attributes depending on the mission.”

The department runs commercial frontier models both in classified systems and in unclassified operational tools such as the new GenAI.mil platform. DefenseScoop reported that suppliers, deployment methods and security levels vary by workspace.

Which court matters?

The legal fight is split in two because the designation rests on two statutes. Mayer Brown says one is 10 U.S.C. § 3252, which requires the Secretary to find in writing that the action is “necessary to protect national security by reducing supply chain risk” and that “less intrusive measures are not reasonably available.” That statute defines supply chain risk as the risk that “an adversary may sabotage, maliciously introduce unwanted function, or otherwise subvert” a covered system. The other is the Federal Acquisition Supply Chain Security Act of 2018.

Anthropic sued on March 9 in the Northern District of California and in the D.C. Circuit. Tech Policy Press reported that on March 26 Judge Rita Lin granted a preliminary injunction, finding in a 43-page ruling that the government’s actions were likely unlawful retaliation. On April 8, a three-judge D.C. Circuit panel denied Anthropic’s request for a stay, citing the risk of forcing the military “to prolong its dealings with an unwanted vendor of critical AI services in the middle of a significant ongoing military conflict.” In August, Lin found the designation violated the First Amendment.

Michael said Anthropic “form-shopped” by filing in California and called the D.C. Circuit case “the more important one.” That court, he said, “goes right between there and the Supreme Court,” and he thinks it is “more of a national security-oriented forum.” “We should hear from them soon, [in] the next month or two. So, there’s still a supply chain risk because that court didn’t designate a preliminary injunction.”

Later, on September 25, the D.C. Circuit upheld the designation 2-1, CNBC reported. Judge Katsas wrote for the majority, Judge Rao joined him and Judge Henderson dissented.

Inside and outside “dot mil”

U.S. officials with access to GenAI.mil described to DefenseScoop the government-tuned frontier models that recently came online there, and asked not to be named so they could speak candidly. “I’ll use it for anything I can,” one said. “It is not perfect and it’s not the end all be all, but I’ve now actually given my team an edict — I said, ‘If you’re going to bring me a product and you haven’t run it through AI at least once, you don’t get to come into my office.’” The point, the official added, is to force people to learn, “because it’s a different skill.”

Another official warned against overselling “where we are in the AI journey” and compared the tools with the model the department is dropping. “I pay for Claude, and I’ll go sit at my computer at home, and in 20 minutes I can get the amount of work done that would take me four hours on government systems.” The official listed what is still missing: “There’s no API access to GenAI.mil. I can’t build agents. I can’t have GenAI.mil inside my inbox.” There is a vision for getting there, the official said, “but we have a lot of work to do to get there.”