Anthropic Won in Court. The Pentagon Is Still Leaving by September 30.

On August 27, a federal judge in San Francisco told the Department of War that it had broken the law when it branded Anthropic a national security threat. Five days later, the department published its refreshed roster of approved generative AI tools for three million military and civilian personnel. Claude was not on it.

Instead, GenAI.mil — the Pentagon's centralized portal for commercial AI — added OpenAI's ChatGPT Mil and Starshield AI's Grok for Government, both accredited to handle Controlled Unclassified Information at Impact Level 5, joining Google's Gemini for Government, which launched the platform in December 2025. The department said more than 1.7 million unique users have signed on so far. And the internal deadline for finishing the removal of Anthropic's products from defense systems has not moved: September 30, 2026.

That is the story the court ruling did not change. Judge Rita F. Lin's decision in Anthropic PBC v. U.S. Department of War (No. 26-cv-01996-RFL) vacated the "supply chain risk" designation that Secretary of War Pete Hegseth imposed under 10 U.S.C. § 3252 — a statute written for sabotage and malicious code, not for a vendor with an inconvenient usage policy. Lin found First Amendment retaliation, a Fifth Amendment due process violation, and arbitrary and capricious agency action. "The empty invocation of national security is not a blank check to punish and retaliate against government critics," she wrote.

What she did not do — could not do — is order anyone to buy anything.

A remedy that stops at the purchase order

Aron C. Beezley, a government contracts attorney at Bradley Arant Boult Cummings who published an analysis of the decision on August 31, put the limit plainly: the court "left standing the government's basic right to choose its own AI vendor," and "the ruling does not require anyone to continue using Anthropic's products."

That distinction is the whole ballgame. Lin's order dismantles the machinery of exclusion — the secondary boycott ordering defense contractors to sever unrelated business with Anthropic, the government-wide bar imposed without notice, the designation itself. It leaves untouched the ordinary discretion of a contracting officer not to renew.

The migration is already far along. In June, Punchbowl News reported that Emil Michael, the Pentagon's chief technology officer and under secretary of war for research and engineering, said at least two-thirds of the department's Anthropic usage had been switched to competing providers. Later reporting has put the figure closer to 90 percent, though The Vault has not independently confirmed that number. An analysis by attorneys at Jones Walker in April noted that even while the Northern District of California injunction was in force, "removal of Claude from DoW systems continues on a 180-day timeline" and contract cancellations proceeded. The Associated Press reported the terminated Anthropic work was valued at an estimated $200 million.

Michael has been the most candid voice on the government side about what actually broke the relationship, and it was not sabotage. In an interview with the Ruthless podcast released August 27 — the same day as Lin's ruling — he recounted summoning the company after reading the inherited contract: "We'll use it lawfully, for all lawful purposes, but it's artificial general intelligence. You don't get to tell me where I could fly a plane to. We decide when we buy a plane from Boeing or Lockheed. And they're like, 'We think these are our moral values. We have our own constitution.'"

Anthropic's response to the ruling was measured. "We welcome the court's ruling that this supply chain risk designation was unlawful," a company spokesperson said in a statement shared with TechCrunch. "We remain focused on working productively with the government to harness AI for our national security so all Americans benefit from this technology." Michael, for his part, wrote on X that the ruling contained "dozens of factual errors" and was issued "during a time of conflict." The government is expected to appeal.

Why this matters: safety policy as a procurement liability

The Anthropic case establishes something useful for every AI vendor selling into government: a usage restriction is protected commercial speech, and an agency cannot use national security authorities to punish you for defending it in public. That is a real legal floor, and Beezley's read is that it will constrain how § 3252 and similar exclusion tools get deployed against difficult vendors going forward.

It is also, on the evidence of the past week, nearly worthless as a commercial defense. The lesson a competitor draws from this year is not that safety restrictions are illegal — they plainly are not. It is that they are expensive, and that the cost lands long before any court gets involved. OpenAI signed a Pentagon arrangement without an Anthropic-style standalone right to prohibit otherwise-lawful government use. xAI's Grok is now in the portal. Google's Gemini has been there since December. Anthropic, which held the incumbent position through its Palantir and AWS integrations, spent six months winning a lawsuit and lost the account.

There is a structural asymmetry here that no injunction fixes. Reversing a designation takes a judge; reversing a migration takes a procurement decision by the same officials who ordered it. Once Claude is out of the software stack, the switching cost runs the other direction, and "we already moved" becomes a facially neutral reason to keep it that way.

The next markers are concrete. The parallel D.C. case remains live, and the D.C. Circuit's April denial of Anthropic's emergency stay means the Federal Acquisition Supply Chain Security Act designation has continued to operate for covered systems even as the California ruling landed. Watch whether the government appeals Lin's order to the Ninth Circuit; whether the September 30 removal date actually holds; and whether Claude reappears on GenAI.mil at all. If the deadline passes quietly and the portal stays a three-vendor shop, the practical answer to whether a company can win this fight will already be on the record.