The Pentagon's AI Portal Goes Live, Still Without Claude

The Pentagon's in-house answer to ChatGPT is now open to nearly everyone who works there. On August 31 the Defense Department folded two more commercial systems, OpenAI's ChatGPT Mil and xAI's Grok for Government, into GenAI.mil, the secure portal it first launched last December on Google's Gemini. The platform is meant to reach all three million of the department's uniformed, civilian and contractor personnel, and officials say more than 1.7 million people have already signed in, according to Military Times and DefenseScoop.

The idea behind GenAI.mil is simple enough. Staff get access to frontier models inside a walled environment rather than pasting sensitive material into consumer apps. Each of the models on the portal has cleared Impact Level 5, the authorization that covers non-public, sensitive unclassified data. For a workforce that has spent two years quietly using outside chatbots without much oversight, that is the point: give people the tools they already want, on infrastructure the department controls.

One name is missing

The list of vendors is notable for who is not on it. Anthropic, maker of Claude and one of the most widely used AI systems in American government, remains shut out. That absence is not an oversight. It is the visible edge of a dispute that TechCrunch has traced back to February, when the department labeled Anthropic a supply-chain risk and dropped it from a set of classified-network deals awarded to rivals including OpenAI, Google, Microsoft and Elon Musk's SpaceX.

The friction is about terms of use. Anthropic wanted contract language ruling out two uses of Claude: mass domestic surveillance and fully autonomous weapons that pick and strike targets without a person in the decision. The department wanted broader wording covering all lawful purposes. Chief executive Dario Amodei has said the company cannot in good conscience sign away those limits. The Pentagon, for its part, argued that Anthropic could meddle with deployed models mid-operation, letting them drift or hiding a backdoor.

A federal judge did not buy that argument. In a ruling we covered last week, the court noted that the models the department already runs are static copies that Anthropic cannot reach, modify or switch off, and called the claimed danger "entirely unfounded." The GenAI.mil rollout shows how little that finding has changed on the ground. Claude is still off the menu, and the standoff over what a military customer is allowed to demand from an AI vendor is far from settled.

A test of the government market

For OpenAI and xAI, a captive audience of three million users is a serious foothold, and it lands as both companies chase federal contracts worth far more than the revenue. For Anthropic, the exclusion is a live cost of drawing a line other labs were willing to blur. The company is heading toward a public listing at a reported valuation near two trillion dollars, and its pitch to investors leans on being the safety-first option. Losing the Pentagon's flagship AI program is the price of that positioning, at least for now.

The bigger question sits underneath the vendor scramble. GenAI.mil is a research and productivity tool, not a weapons system, and nobody is claiming otherwise. But the argument that kept Claude out, over who sets the rules on lethal autonomy and surveillance, is the one that will matter most as these systems move closer to operational work.