Pentagon Adds ChatGPT and Grok to Its AI Portal
The Pentagon's in-house chatbot is no longer just a Google product. Officials confirmed Monday that anyone signing into GenAI.mil — the department's enterprise AI platform — can now choose between approved versions of OpenAI's ChatGPT, Starshield AI's Grok, and Google's Gemini, which had the place to itself since launch.
That finishes a rollout promised back in December 2025, when the department said it would put GenAI.mil in front of more than 3 million service members, civilian employees and contractors. Gemini for Government went first. ChatGPT and Grok were always supposed to follow.
What's actually on the platform
Both new models cleared Impact Level 5, the highest authorization the Pentagon grants for systems handling non-public, sensitive unclassified data. That's the gate that matters — it decides which work can move onto the platform at all.
ChatGPT Mil, per the department's announcement, brings a "familiar commercial experience" behind the secure wall. The work it's pointed at is document-heavy and deeply unglamorous: planning, policy, logistics, administration. Officials described the core as chat, files, projects and custom GPTs, "with additional features sequenced over time."
Someone familiar with the work at OpenAI put it more plainly. The day-to-day use is repeatable paperwork — travel forms, responses to contracting opportunities. Staff have GPT-5.4 Terra now and will get 5.6 Terra soon. There's also an offline search function that returns citations and can be refreshed as new information lands.
Grok for Government — xAI's federal model, now sold through Starshield AI after Musk merged xAI and SpaceX earlier this year — leads with reasoning. Deep-thinking inference, three modes called Auto, Fast and Expert, customizable workspaces, persistent projects, and reusable "playbooks" meant to capture how a team does something and spread it. xAI and SpaceX didn't respond to requests for comment.
The vendor-lock argument, and the model that's missing
The department pitched all this as insurance against leaning on one company. A source who asked to be identified as a War Department official — a title that comes from an executive order signed last year, though the actual renaming would take an act of Congress — said the department "will continue to build an architecture that prevents AI vendor lock and ensures long-term flexibility for the Joint Force."
Which is a strange thing to say while one of the biggest providers sits outside the building.
Anthropic was on the original list. Its leadership wanted contract terms barring the department from using its models for mass surveillance of Americans or in connection with lethal autonomous weapons. The Pentagon wanted unrestricted access for whatever it deems lawful. About six months ago it designated Anthropic a supply-chain risk and ordered the company's products pulled from national security systems. Anthropic sued. Last week a federal district judge ruled the designation and everything that followed were illegal and baseless.
So the portal now advertising model diversity is short one major model, over a fight the courts have started to settle.
Why it matters
Three million people is a staggering rollout, and nobody is pretending otherwise. "We are moving at wartime speed: empowering the department to experiment, iterate, and deploy capabilities faster than our adversaries can adapt," the official said.
Speed cuts both ways. These models write with total confidence whether or not they're right, and that habit doesn't improve because the server is behind a firewall. Impact Level 5 governs where the data sits. It says nothing about whether the summary at the bottom of a logistics memo is accurate — and a travel form is a low-stakes place to find that out, while a planning document is not.
Two things worth watching: whether last week's ruling reopens the door for Claude, and whether the department ever publishes real usage numbers once the novelty wears off.
Image: Ivan Chumak, via Pexels
