D.C. Circuit leaves the Pentagon's Anthropic ban in place for contractors' department work

The two rulings sit on separate legal tracks. According to Reuters, the cases proceed through different legal routes and address government actions under distinct claims ^[5]. The California judge granted Anthropic summary judgment on several claims, First Amendment and due process among them. The order describes that case as covering a broader set of government actions ^[3]. The Washington case is Anthropic's direct challenge to the designation ^[6]. The designation followed a dispute over whether the department could require uses of Claude that Anthropic had excluded from its contracts ^[17].

The reports disagree on what kind of decision came out of Washington. Reuters describes an interim step: the court declined to block the Pentagon's action, and that decision does not by itself resolve whether the designation was lawful ^[18]. CNBC, as summarized by The Decoder, reported something closer to a merits holding, a 2-1 ruling that the Pentagon was right to classify Anthropic as a risk ^[2].

For a contractor, the scope that applies is in the D.C. Circuit's April order, and it follows contract lines. The Pentagon canceled its own Anthropic contracts and barred other contractors from using Anthropic as a subcontractor on work performed for the department ^[12]. The same order said the restriction did not stop those contractors from using Claude on work for other customers ^[13].

The design I would defend at review is small. Model choice goes behind one interface keyed on contract ID. A second model gets its own prompts and an eval suite built from department workloads, because a prompt tuned for Claude has only been tested on Claude. Department routes default to that model now. Every request logs the contract and the model that served it.

In my view, the public positions leave little room for a settlement. Amodei argued in February that current AI systems are not reliable enough to select and engage targets without human involvement ^[9]. Defense Secretary Pete Hegseth argued, according to The Decoder, that Anthropic's safety restrictions could jeopardize military operations ^[10]. Anthropic rejects the Washington ruling and says it is weighing its next steps ^[7].

Intelligence customers are the least clear case. Anthropic said Claude was already deployed for national-security work including intelligence analysis, operational planning and cyber operations ^[15]. The Decoder describes US intelligence agencies as heavy users of Anthropic's models and calls the fit with a supply-chain-risk label an open question ^[16]. The April order, as reported, is written around work performed for the department ^[12].