Three things happened in the two weeks before oral argument in Anthropic PBC v. U.S. Department of War, No. 26-1049 (D.C. Cir.), set for May 19. Each is, on its own, an item of public record. Taken together, they reshape parts of an evolving D.C. Circuit court record that is meaningfully different from what existed when that Court issued its stay denial on April 8.

The Record as of May 1

On April 17, Anthropic CEO Dario Amodei met at the White House with Chief of Staff Susie Wiles and Treasury Secretary Scott Bessent. The meeting was described as “introductory” and “constructive,” and focused on Anthropic’s new Mythos model.

What This Means for the Upcoming Argument

The D.C. Circuit's April 8 order did not reach the merits. Rather, three developments alter the Nken-based decision balance in two ways — ways that could have an impact on the scheduled May 19 appellate argument.

First, the panel weighed the government's interest against Secretary Hegseth's January 9 directive that the Department "must also utilize models free from usage policy constraints," and against a record showing that the United States is "in the middle of a significant ongoing military conflict." It framed the harm of a stay as "judicial management of how, and through whom, the Department of War secures vital AI technology" during that conflict. The May 1 announcement, however, confirms that seven major vendors are already contracted to deploy on classified networks at high security levels. The question is whether enjoining a single designation still carries the same weight, given the Department’s own representation of alternative supply.

Second, on the protected-speech side of the balance, the panel discounted Anthropic's First Amendment framing and observed that Anthropic had not shown its speech was actually chilled during the litigation. The April 30 hearing testimony is now an artifact in the public record. Whether it bears on the First Amendment retaliation analysis Judge Rita Lin accepted in the parallel Northern District of California proceeding (now on appeal to the Ninth Circuit) is a question for the panel and the parties to address; courts considering retaliation claims under Nieves v. Bartlett, 587 U.S. 391 (2019) and Hartman v. Moore, 547 U.S. 250 (2006) have considered contemporaneous statements by responsible officials when they bear on the action at issue. How much weight that statement carries is for the panel to decide.

A third point bears mention. The reported April 17 meeting between Mr. Amodei, Ms. Wiles, and Secretary Bessent describes one part of the executive branch transacting with Anthropic at senior levels, while another part publicly designates the company a supply-chain risk. The doctrine of unconstitutional conditions, articulated in Agency for International Development v. Alliance for Open Society International, 570 U.S. 205 (2013), is most useful where the government's coercive instrument is targeted rather than generalized. The April 17 event doesn't necessarily establish that proposition, as named-source confirmation is limited, and the meeting is described as introductory. But this may be reconciled with the asserted irreparability of dealings with the same vendor.

"AI-First" and the Question of Architectural Safety

The May 1 announcement was procurement news. It was also the deployment side of an analytical question we previously flagged in Part 3. Frontier AI capabilities deployed on the Department's networks have downstream paths that, in our prior posts, we traced from defense systems into federal and local law enforcement-adjacent platforms. The Department's framing is "decision superiority across all domains of warfare." The procurement question is whether the safety architectures the contracting vendors apply on commercial deployments will translate to classified deployments, and the extent to which the categorical constraints that mattered in Carpenter v. United States, 585 U.S. 296 (2018), and that the European Union codified in Article 5 of the AI Act will carry.

The May 1 announcement may not vindicate any specific safety choice Anthropic made. The post-delivery technical access question the D.C. Circuit asked the parties to brief — "whether, and if so how, Anthropic is able to affect the functioning of its artificial-intelligence models before or after the models … are delivered to the Department" — is identical in form for every contracting vendor. What changed on May 1 is the scale and breadth of the deployment record against which that technical question is to be answered, and the corresponding need for contracting officers and oversight counsel to ask it of the seven prospective vendors.

Canary or Outlier?

The May 1 announcement begs the question of whether Anthropic's pre-litigation refusal to drop its usage policies was a vendor-specific posture (the "outlier" reading) or an industry-wide structural diagnosis that the courts and Congress will eventually revisit (the "canary" reading). Both readings are coherent on the current record.

The honest analytic move is to specify what would distinguish them. If the seven contracting vendors are deploying with materially weaker architectural constraints on classified networks than they apply to commercial customers, the canary reading gains empirical support. If their architectures impose comparable constraints through different mechanisms — RLHF tuning, system-prompt hardening, output-class filters, contractual usage restrictions — then Anthropic's posture is one defensible architectural choice among several, with a different surface but a similar interior. We don't yet have the documents to pick. The Department's announcement does not specify the deployment posture of any contracting vendor with the granularity the question requires, and the contracts themselves are not public.

The point here is not to choose between the two readings. It is to identify the test that will eventually distinguish them, and to flag that the ongoing D.C. Circuit direct review proceeding and the parallel Ninth Circuit appeal are the most likely venues for development of the relevant facts.

Anthropic PBC v. U.S. Department of War, No. 26-1049 (D.C. Cir.); Anthropic PBC v. U.S. Department of War, No. 3:26-cv-01996-RFL (N.D. Cal.).

---

For questions about AI governance, government technology procurement, First Amendment implications of regulatory action, and vendor risk management, contact the Jones Walker Privacy, Data Strategy, and Artificial Intelligence team. And stay tuned (and subscribe) for continued insights from the AI Law and Policy Navigator.