i
DATAIST
News · 2026-08-31

Pentagon adds ChatGPT Mil and Grok to GenAI.mil, but not Claude

@neuronium_ai @neuronium_ai

The Pentagon has put ChatGPT Mil and Grok for Government inside GenAI.mil, the centralised secure portal it launched last year so Department of Defense staff could work with commercial frontier models without sending sensitive government information into ordinary consumer services. Google Gemini was the only model available when the portal opened. Of roughly 3 million defense department employees, more than 1.7 million unique users have connected to it, according to the department. Anthropic's Claude is not on the menu.

Cover: Pentagon adds ChatGPT Mil and Grok to GenAI.mil, but not Claude

The Pentagon has put ChatGPT Mil and Grok for Government inside GenAI.mil, the centralised secure portal it launched last year so Department of Defense staff could work with commercial frontier models without sending sensitive government information into ordinary consumer services. Google Gemini was the only model available when the portal opened. Of roughly 3 million defense department employees, more than 1.7 million unique users have connected to it, according to the department. Anthropic's Claude is not on the menu.

ChatGPT Mil comes through OpenAI for Government and will look familiar to anyone who uses the commercial product. The department lists four capabilities: chat for talking to the model, files for working with documents, projects for organising tasks, and custom GPTs for building specialised assistants. It is meant for routine document work and other unclassified jobs — administrative processes, logistics, planning, and drafting policy papers. More features are promised later.

The department describes Grok differently, and more expansively. Grok from Starshield AI, it says, should give the military a quick productivity gain, help preserve working knowledge, and make collaboration safer and more efficient. Starshield AI is SpaceX's secure satellite network, built on existing Starlink technology. The department also says Grok will let service members complete tasks faster and more accurately across varied operating conditions, offering market analysis for acquisition specialists and supply chain management for logisticians as examples.

Read the two descriptions side by side and the asymmetry is hard to miss. One is an office suite: chat, files, projects. The other is written in the vocabulary of operations — operating conditions, knowledge retention, safer collaboration — and arrives attached to a satellite network. Both sit behind the same unclassified login. The difference so far is in how the buyer is describing them, not in what has been deployed.

The military versions, the department says, also stay out of the data collection that is difficult to decline in consumer technology products.

The Pentagon has signed agreements with several other technology companies to broaden its AI capabilities: Amazon Web Services, Microsoft, Nvidia and Reflection AI.

Claude's absence is the part of this worth slowing down over. The Trump administration declared Anthropic a supply chain threat after the company refused to give the Pentagon unrestricted access to its AI tools and insisted on keeping a set of protective limits in place. Anthropic is challenging that designation in court. Whatever the litigation decides, the commercial outcome is already visible in this announcement: a frontier lab that held a line on usage restrictions lost a distribution channel that reaches more than half of a 3 million-person department, and two labs that did not hold that line now occupy it.

That 1.7 million figure is what makes this a procurement story rather than a product story. A portal that most of the defense department has already touched turns the question of which models sit behind the login into a distribution decision with an unusually large denominator. Vendors do not typically get a second chance at that kind of placement.

Notably absent from the department's description is any statement of what OpenAI and xAI agreed to. The Pentagon's grievance with Anthropic is specific and public — it wanted unrestricted access and did not get it. Nothing published about ChatGPT Mil or Grok for Government says what restrictions, if any, the other two accepted, which is the single fact that would make the comparison legible. We know what the department said no to. We do not know what it said yes to.

A supply chain threat designation is normally reserved for a supplier that might fail to deliver. Anthropic's offense was delivering with conditions attached. If that reading survives court, the terms on which frontier models reach the US military stop being negotiated and start being dictated — and the next lab weighing whether to keep its guardrails will have this portal, and its 1.7 million users, sitting on the other side of the scale.