Pentagon Says It Has Moved 90% of Classified AI Work Off Anthropic

The Pentagon wanted AI software it could use for all lawful purposes without contractual limits. It now says it has shifted about 90% of its classified AI workloads away from Anthropic and expects to finish the transition by the end of September.

The update, from Under Secretary of Defense for Research and Engineering Emil Michael, puts a date on a rapid break that had already reached deeply into military AI systems. Michael said all Maven Smart Systems or Palantir work had moved away from Anthropic months earlier, leaving the remaining classified workload as the final part of the shift.

A break over control

The transition followed a breakdown in negotiations over how Anthropic’s models could be used. Anthropic sought contract safeguards restricting uses connected to mass surveillance of U.S. citizens and fully autonomous lethal weapons systems. The Defense Department rejected those terms, arguing that defense software needed to be available for all lawful purposes in an unconstrained way, and later designated Anthropic a national-security supply-chain risk.

About 90%Classified AI workloads transitioned

Emil Michael said about 90% of classified AI workloads had moved away from Anthropic models.

End of September 2026Target for remaining workload

The Defense Department said it was on track to complete the transition by the end of September.

From one supplier to several

Michael has framed the move as more than a replacement exercise. The Pentagon is pursuing a multi-model system that uses different frontier AI vendors across missions and security environments. DefenseScoop reported that the mix includes OpenAI, Google and xAI, while the suppliers, deployment methods and security levels can differ by workspace.

Michael’s argument is that relying on one model creates its own military risk. If a sole supplier falls behind the frontier, he said, warfighters could be left in a suboptimal position. A broader vendor mix, in that view, lets the department choose models with different attributes as the field changes.

The operational shift does not end the dispute

The migration is proceeding while the underlying conflict remains in court. Michael said a D.C. Circuit case involving the supply-chain-risk designation was still pending, and he expected movement within a month or two. The legal process therefore remains separate from the department’s immediate effort to remove Anthropic from classified work.

That separation is the central consequence of the September deadline. The Pentagon is trying to build a classified AI environment that does not depend on a provider willing to impose use restrictions, while Anthropic’s challenge to the government’s response continues on a different track.