Court reversal turns the Pentagon's Anthropic ban into a migration teams may have to run twice

Federal contractors have spent part of the spring pulling Claude out of document pipelines to meet agency deadlines. Trump's late-February order gave agencies six months to get off Anthropic's products ^[7], which puts the deadline somewhere near the end of August ^[13]. That work does not come back. Judge Rita Lin's decision stops certain federal agencies from enforcing the order ^[11] and says the supply chain risk label was illegal and baseless ^[1], but it does not require the Pentagon to resume working with the company ^[9].

The reasoning behind the label reads more like a press response than a procurement decision. Lin wrote that "the empty invocation of national security is not a blank check to punish and retaliate against government critics" ^[4], and that neither the Constitution nor the statute the government invoked permits sweeping penalties imposed principally because of Anthropic's critique of the administration's views ^[5]. According to Engadget's account of CNBC's reporting, the trigger was the Defense Department and Secretary Pete Hegseth pressing Anthropic to remove certain safeguards, with Dario Amodei refusing uses tied to mass surveillance and autonomous weapons ^[10][2][3]. That chain is a dispute about criticism and retaliation. It has nothing to do with code, uptime, or delivery.

The awkward part for buyers is that Anthropic won one of two cases. It sued over two separate designations, and it remains designated a supply chain risk while the second decision is pending ^[8]. A procurement officer therefore holds a vendor that is both cleared and blacklisted, depending which docket you read.

The commercial signal points the other way. Anthropic reported record revenue of $11.5 billion, which Engadget describes as 14 times the same quarter a year earlier ^[14]. Divide and the year-ago figure lands around $820 million ^[16], and the company is still headed for a likely near-record IPO this year ^[15]. The risk here is transactability, not vendor survival, and it fails differently: the vendor can be healthy while your contract is void.

The useful exercise for a dependency review is three numbers per model provider. What it costs to migrate off. What it costs to migrate back. How many working days you would actually get between an order and a compliance date. Six months sounds generous until it has to cover both the removal and the re-scoring of every eval you tuned against one model's refusal behavior. If off-plus-back exceeds the cost of keeping a second provider warm through your release cycle, keep it warm. The tradeoff is not free: a warm second path means duplicated prompt work and two eval suites maintained every release, for a scenario that may never fire.

This is for teams whose revenue passes through a federal customer, and for anyone whose provider has published limits it intends to defend under pressure ^[3]. Anthropic's refusal is the sort of commitment a buyer might have chosen it for, and it is also what put that buyer into a migration. The court has now shown that the resulting ban can be undone in litigation ^[1][11].