Pentagon Says It Stopped Using Anthropic's Claude: What It Means

On October 5, 2026, the BBC reported that the U.S. Department of Defense has stopped using Anthropic's AI tools, citing a department official who said the Pentagon had ceased use of Anthropic products. The statement marks the formal end of a phaseout ordered after Defense Secretary Pete Hegseth designated the company a national-security supply-chain risk in February. The same BBC report adds a complication: multiple people familiar with the matter told the broadcaster that Claude was still being used as recently as the week before, including in research, analysis and intelligence work and in operations against Iran, through Palantir's Maven Smart System. This article sets out what is confirmed, what is disputed, and what it means to "stop using" a model that is embedded inside other vendors' software.

Anthropic

Palantir

OpenAI

Google

xAI

The vendors at the center of the story: the model supplier being removed, the integrator that embedded it, and the labs reported to be replacing it.

Feb 27, 2026Supply-chain-risk announcementBy Defense Secretary Hegseth

6 monthsOrdered phaseout windowTarget was late August

2 to 1D.C. Circuit voteUpheld the designation, Sep 25

$200MCeiling per lab contractFour July 2025 deals, per The Intercept

The Pentagon stops using Anthropic's Claude: what was said

The Defense Department's account is short: the phaseout is complete. According to coverage of the BBC report, an official said the department has ceased the use of Anthropic products. The BBC piece also quotes Lauren Kahn of Georgetown's Center for Security and Emerging Technology, who said that the length of time it took to remove Claude from across the system "is indicative that these things are not just plug and play."

The second account comes from the BBC's unnamed sources, who say Claude remained in use days before the announcement. These two statements are not necessarily contradictory. A department can truthfully say a phaseout finished on a given date, and sources can truthfully say a tool was running a week earlier. What the gap shows is that the original deadline, reported as late August, was not met, and that the end date was reached some weeks later. Coverage also relays comments from Emil Michael, the Under Secretary of Defense for Research and Engineering, on September 11 that "about 90% has transitioned" and that Maven and Palantir work had moved months earlier. The BBC's sources describe a different picture for Maven, and the public record does not resolve it.

Claude phaseout: planned window vs reported milestones

Days elapsed since the February 27, 2026 supply-chain-risk announcement. The grey bar is the planned six-month window (approximate); green bars are reported milestones.

Sources: BBC News (Oct 5, 2026) as relayed by AI Weekly and The Star, ABC News (Sep 25, 2026). Day counts are our own arithmetic from reported dates. Hover a bar for details.

Timeline of the dispute

The sequence matters, because the legal and operational tracks moved at different speeds.

February 26 to 27, 2026. Anthropic CEO Dario Amodei said the company "cannot in good conscience accede" to a contract term allowing "any lawful use." Anthropic's stated exceptions were mass domestic surveillance and fully autonomous weapons. The next day Hegseth announced the supply-chain-risk designation, with a formal determination reported on March 3.
March 2026. Fortune reported that Michael described a realization of dependency, asking what would happen if the software went down, and later said "I need redundancy." Anthropic sued.
August 28, 2026. A federal judge in San Francisco, Rita Lin, ruled one designation unlawful retaliation. We covered that decision in our analysis of the ruling.
September 25, 2026. The D.C. Circuit, 2 to 1, upheld a parallel designation. Judges Gregory Katsas and Neomi Rao formed the majority, and Judge Karen LeCraft Henderson dissented. ABC News reported the majority found the department had enough evidence to conclude that Claude's built-in restrictions and the unresolved contract dispute could make it unreliable for military operations.
October 5, 2026. The department tells the BBC that use has ceased.

The two court outcomes point in different directions because they concern two separate designations, and The Next Web reports that Anthropic said it remains confident and is considering all options, including further review. Hegseth said the appeals ruling "completely validates the Department's position."

The fact that it took the Pentagon until now to remove all use of Claude from across the system is indicative that these things are not just plug and play.
Lauren Kahn, Georgetown CSET, quoted by the BBC

What "stopped using" means operationally for an embedded model

A model like Claude is rarely a standalone application inside a large defense system. It is more often a component called by other software. That changes what removal involves, and the points below are our analysis of the general pattern rather than reporting on specifics the Pentagon has not disclosed.

Contracts versus code paths. Ending a direct contract is a paperwork action. Removing a model from a deployed workflow means finding every place a prompt is sent to it, replacing the endpoint, and re-testing outputs. Prompts tuned for one model often behave differently on another, so each substitution needs validation, not just a configuration change.

Integrators carry the dependency. According to the BBC, Claude was embedded in Palantir's Maven Smart System, the platform that fuses satellite imagery, drone footage and other data for intelligence and targeting work. When a model sits inside an integrator's product, the customer's relationship runs through the integrator. A supply-chain designation that bars contractors from using a model, as The Next Web describes, therefore pushes the change down into the prime contractor's own engineering schedule. Maven's scale makes that nontrivial: DefenseScoop reports that its user base more than doubled from 50,000 in January to over 100,000 following Operation Epic Fury, and that Maven was supporting strikes on 13,000 targets in 38 days.

Classified environments move slowly. Models approved for classified networks go through accreditation, and replacements need the same. The public sources reviewed here do not detail the accreditation status of the replacement models, so how much of the transition rested on already-cleared alternatives is not established.

"Stopped" can mean several different things. It can mean no new tasks, no active contracts, no calls to the API, or no remaining dependence anywhere in the supply chain, including subcontractor tooling. The official statement as reported does not specify which definition applies, and that ambiguity likely explains how both the department's claim and the sources' account can coexist.

Aerial view of the Pentagon and surrounding roads in Arlington, Virginia, during construction work — An aerial view of the Pentagon from a U.S. Navy photograph filed in 2003, showing renovation work near the river entrance. The photo predates the AI contracts discussed here and is shown only to illustrate the building. Public domain, via Wikimedia Commons.

Vendor concentration in defense AI

The dispute exposed how much a single supplier had become part of the department's tooling. Fortune's March reporting quotes Michael describing the moment he recognized the dependency, and the department's response was to broaden its supplier base. Per the BBC, the Pentagon has since signed agreements with Google, xAI and OpenAI. That fits the pattern we described in our look at GenAI.mil, where ChatGPT and Grok joined Gemini behind one portal, and in our coverage of the FOIA-released contracts, which showed four labs each holding a deal worth up to $200 million.

Two kinds of concentration are worth separating. The first is model concentration: relying on one lab. The second is integrator concentration: relying on one platform, such as Maven, through which models reach users. Maven itself has grown, with DefenseScoop reporting its contract ceiling rose from $480 million to nearly $1.3 billion through 2029 in May 2025. Diversifying across model labs reduces the first risk but does not by itself address the second, since one integrator can still be the single path to many users.

The episode also shows a different risk from outages or price changes: a vendor's usage policy becoming a procurement variable. The D.C. Circuit majority treated a lab's built-in restrictions as a legitimate reliability concern for the military buyer, while Judge Henderson's dissent argued the statute does not cover a contractor's open enforcement of stated restrictions. Courts have so far not produced one consistent answer, and buyers in any sector can reasonably read that as a reason to keep their model layer swappable. Model-agnostic workspaces such as Metir are built around that premise, though the defense setting is far more constrained than most commercial ones.

What to watch next

Whether Anthropic seeks further review of the D.C. Circuit decision, and how the two court tracks are reconciled.
Whether the department or Palantir publishes details on how Maven's model layer was changed, and how transition was validated.
Whether independent reporting confirms or refutes the BBC sources' account of use in the final week.
How replacement agreements with Google, xAI and OpenAI are structured, including usage terms, since the question of "any lawful use" language was the core of the dispute.

Sources:

Image credits

Hero: the Pentagon, DoD photo by Master Sgt. Ken Hammond, U.S. Air Force, via Wikimedia Commons, public domain (U.S. federal government work). The photograph is undated.
In-body: aerial view of the Pentagon river entrance, U.S. Navy photo (file dated 2003), via Wikimedia Commons, public domain (U.S. federal government work).