The US military's in-house AI platform just got a lot bigger, and the timing tells a story of its own.

On Monday the Department of Defense said OpenAI's ChatGPT and xAI's Grok had joined Google's Gemini on GenAI.mil, the Pentagon's enterprise generative-AI portal. The platform is being rolled out to roughly 3 million military, civil-service and contractor personnel — about 1.5 million were already using it as of June — and it handles non-public, sensitive-but-unclassified data at the Impact Level 5 tier. OpenAI's entry is a defence-tuned build called ChatGPT Mil; Grok arrives through xAI's government unit, Starshield AI, as "Grok for Government." Officials framed the expansion as choice. "Access to a diverse suite of AI capabilities ... will give warfighters the tools they need," a defence official said, and OpenAI's head of government, Joe Larson, called the integration "grounded in our national security principles."

The lab that isn't there

One frontier name is conspicuously missing: Anthropic, and its model Claude. That absence is not an oversight. Anthropic was, in fact, the first AI company to put its models on the Pentagon's classified systems — before the relationship broke down and the department formally designated it a "supply-chain risk" roughly six months ago, ordering Claude purged from its systems.

The dispute was about terms, not technology. Anthropic wanted contractual guardrails: that its models not be wired into lethal autonomous weapons, and not be turned on Americans for mass surveillance. The Pentagon wanted the model without those conditions. When the two sides could not agree, the "supply-chain risk" label followed — a designation with the practical effect of freezing a vendor out of federal work.

A court win that hasn't changed the platform

Here is where the timing bites. Just last week, US District Judge Rita Lin, in the Northern District of California, vacated that designation as unconstitutional, finding it breached the First and Fifth Amendments. (We covered the original dispute and the ruling as they landed.)

And yet, days after that decision, the Pentagon's flagship AI platform added two of Anthropic's rivals for millions of users — and Claude still is not among them. A ruling that a blacklist was illegal is not the same as an order to reinstate a vendor, and procurement does not turn on a dime. But the optics are stark: the one lab that fought the government over how its AI could be used, and won, is watching the models that did not draw the same public red lines get wired into the daily work of the US military.

Why it matters beyond one vendor

It would be easy to read this as a corporate spat over a lucrative contract. The more durable question is the one Anthropic's red lines actually raised: on what terms should a commercial AI model be used inside a defence establishment at all? The guardrails at the centre of the fight — no autonomous lethal targeting, no domestic mass surveillance — are exactly the limits that decide whether "AI for the warfighter" stays on the right side of the lines most democracies say they want to hold.

GenAI.mil is, for now, a sensitive-but-unclassified productivity tool for planning, logistics and paperwork, not a targeting system. But it is also the on-ramp: the platform where three million people form their habits about what these models are for. Which labs get to be on that on-ramp, and under what conditions, is a policy choice the Pentagon is making one integration at a time — and the lab that held out over those terms is, for now, the one left off it.