Two federal courts split on the Pentagon's Claude ban by ruling under different laws

A defense contractor that read the August ruling as permission to switch Claude back on was reading half the case. The D.C. Circuit answered a different question, under a different statute ^[3].

The panel's question was whether FASCSA, a 2018 law that lets the Defense Secretary block a supplier whose technology poses a national security risk, allowed the ban ^[7]. That law requires the Pentagon to notify a supplier and give its reasons before blocking it. Anthropic argued the department had not completed that process, and the court did not accept the argument ^[10]. The 2011 law the California judge applied sets different requirements for supplier bans ^[12]. "Another federal court has already held the government's parallel designation unlawful," Anthropic said ^[11]. By Anthropic's own description, then, there are two designations under two laws, and beating one leaves the other in force. The report does not say whether contractors can reach Claude today.

Before the ban, Anthropic held a $200 million contract to give the Pentagon access to Claude, and the two sides had discussed offering it through a platform called GenAI.mil ^[4]. The designation the panel upheld was motivated by Anthropic's limits on how Claude may be used ^[8]. "On more than one occasion, these restrictions have stopped Claude from performing tasks requested by government users," the ruling says ^[9]. As the court describes it, the national security risk is refusal behavior Anthropic built in on purpose.

Both rulings arrived within about three months of the June ban ^[14]. Anthropic has not yet picked its next route. "We remain confident in our position and are considering all options, including further review," the company said ^[11].

Teams like to tell themselves the model is one API call they can swap in an afternoon. What users actually depend on is the prompts and evaluations built around one model's behavior, refusals included. For defense-bound work, I think the right move is to integrate a second model now and test it on the same tasks, without waiting for the appeals. The cost is a second set of prompts and evaluations to maintain, and the team pays it even if Anthropic wins.

Two facts about a team's own setup sort the decision. The first is whether the workload reaches the Pentagon or a defense contract. The second is whether swapping the model takes days or a quarter. The June ban hit hardest the teams that are defense-bound and slow to swap, and they should build the fallback first. A defense-bound team that can swap quickly needs the second model kept current, run against real tasks each month and sitting behind a configuration switch. As reported, the ban covers Pentagon staff and defense contractors ^[6]. A team outside that work with a slow swap can follow the rehearing filings and wait. For a team outside defense work that can swap quickly, nothing changes.