The Defense Department has signed AI procurement deals with seven companies for use on classified networks: Amazon Web Services, Google, Microsoft, Nvidia, OpenAI, SpaceX, and Reflection AI. The contracts feed GenAI.mil, an internal Pentagon portal serving more than 1.3 million Defense Department users. Anthropic, which previously held a $200M DOD ceiling agreement, is not on the list โ and that exclusion is the real story.
The dispute is concrete. Anthropic refused to permit "all lawful" Pentagon uses of Claude, specifically blocking applications in autonomous weapons and domestic mass surveillance. Defense Secretary Pete Hegseth designated Anthropic a "supply-chain risk" in March 2026; Anthropic sued and won an injunction against the designation. The litigation is open. In parallel, the Pentagon reportedly continues to use Anthropic's Claude Mythos Preview model โ meaning the policy fight and the operational reality have already diverged. Reflection AI, the seventh winner, is a 2024 startup founded by former Google DeepMind researchers that has not yet announced a commercial product but is expected to ship a language model trained on tens of trillions of tokens.
For builders, the ecosystem signal is bigger than the procurement headline. The DOD is explicitly architecting against vendor lock-in โ seven vendors instead of consolidating around two or three frontier labs. That's good for the open-vs-closed dynamic; less good if you assumed your stack would converge on one or two providers. The Anthropic exclusion also draws a sharper line on usage policy as a market-shaping force: the lab that holds firmest on weapons and surveillance restrictions is the lab that loses the contract, and the lab that backfills is whoever holds firmest on revenue. Whether that's principled or politically expensive depends on which side of the wire you're sitting on.
If you're building on Claude in a government or government-adjacent context, the operational situation is unstable: the Mythos Preview reportedly being used despite the formal exclusion suggests the boundary is being negotiated case-by-case. If you're building on the seven cleared vendors, GenAI.mil is now a real procurement surface โ worth understanding the integration path even if you're nowhere near classified work today, because federal procurement gravity tends to pull civilian standards behind it.