A divided appeals court let the Pentagon exclude Anthropic, and single-vendor model risk stopped being hypothetical

On September 25, a divided panel of the US Court of Appeals for the D.C. Circuit upheld the Pentagon’s designation of Anthropic as a supply-chain risk, clearing the Department of Defense to strip Claude from its systems and to keep defense contractors from using Anthropic’s products on Department work. It is a narrow procurement ruling on its face, but it is also the clearest signal yet that standardizing on a single frontier-model vendor carries a risk most engineering plans do not account for: the vendor becoming unavailable for an entire class of work by government action.

What the court decided

The panel ruled 2-1 on the merits of Anthropic’s petition for review, not on a temporary motion. Judges Gregory Katsas and Neomi Rao formed the majority, with Judge Karen Henderson dissenting. The designation runs under the Federal Acquisition Supply Chain Security Act, the 2018 law that lets agencies exclude a supplier from federal procurement when they judge it a risk.

The dispute underneath the legal question is about how Claude is built. Anthropic had a defense contract in 2025 but negotiations broke down when the Department sought access for all lawful uses and Anthropic declined to remove restrictions meant to keep its models out of fully autonomous weapons and domestic mass surveillance. Writing for the majority, Katsas held that the Department “reasonably feared that Anthropic might manipulate Claude’s design to prevent it from performing national-security functions that the Department deems contractually authorized and necessary.” That reasoning is the part worth sitting with: a buyer excluded a model not for a defect but for the guardrails its maker chose to encode.

The split that keeps it unresolved

This is not the whole story, and the article that told you otherwise would be overstating it. The Pentagon pursued Anthropic along more than one track, and the tracks have split. A federal judge in the Northern District of California, Rita Lin, earlier found that the government acted unlawfully in a parallel action, treating it as retaliation for Anthropic’s public criticism and blocking a broader directive that would have pushed all agencies off Claude. So the D.C. Circuit decision hardens the Defense-specific exclusion while a California injunction still protects Anthropic’s standing elsewhere in government.

Anthropic said it “respectfully disagrees” with the D.C. Circuit and pointed to the California ruling that “another federal court has already held the government’s parallel designation unlawful,” adding that it is weighing further review. A full-court rehearing or a Supreme Court petition is plausible, which means the bottom line for Defense work is set for now but the wider question is genuinely open.

Why a procurement fight matters to builders

For most working developers, a Pentagon contract dispute is background noise. For anyone making platform decisions, it is a concrete data point that changes the shape of “vendor risk.” Until now, the case for abstracting model access behind a gateway rested on price swings, rate limits, and the occasional deprecation. This adds a category that is harder to route around: a model provider can be ruled out of a whole domain of work by a third party, on grounds that have nothing to do with uptime or quality.

The practical response is the one that was already good hygiene and is now easier to justify. Keep model choice behind an abstraction rather than hard-coding one provider’s SDK through the codebase. Route through a gateway that can fail over to another model. Know, before you need it, what switching Claude for a comparable model would actually cost in prompts, evals, and tool integrations. None of this is about betting against Anthropic. It is about not writing code that assumes any single model will always be an option.

What to watch

The immediate signal is procedural: whether Anthropic seeks rehearing or the Supreme Court, and whether the California and D.C. tracks stay contradictory or get reconciled. The larger signal is precedential. If a designation can stand because of how a vendor constrains its model, other large buyers and regulators now have a template, and the safety choices a lab makes become a procurement variable rather than a purely technical one. For teams, the takeaway is not which side is right. It is that model portability has quietly moved from a nice-to-have into a form of insurance.

Sources cited

1. Anthropic PBC v. United States Department of War, No. 26-1049, D.C. Circuit docket and opinion (CourtListener) www.courtlistener.com
2. US appeals court upholds Pentagon's blacklisting of Anthropic, with vote, panel, and majority quote (Military Times) www.militarytimes.com
3. Federal court says Pentagon can label Anthropic a supply chain risk (Broadband Breakfast) broadbandbreakfast.com
4. Two courts, two postures: the dual-track litigation and statutory framework explained (Jones Walker LLP, AI Law Blog) www.joneswalker.com