The Department of War picked two more chatbots for its internal AI platform on October 12. The one it left out is the one it has been arguing with.
ChatGPT Mil, OpenAI’s government offering, and Grok for Government, xAI’s, will join GenAI.mil, the department’s enterprise generative AI platform. Both cleared Impact Level 5, the certification that allows a system to handle controlled unclassified information. Until now, the platform carried only a military version of Google’s Gemini.
GenAI.mil has been live for nine months and has accumulated 1.7 million unique users, by the department’s count. That figure is the part of the announcement that matters as much as the vendor names. A platform with that many users is no longer a pilot. It is a working tool inside an institution that has to move quickly and cannot easily walk away from a system its people already use.
The stated reason for adding two vendors to a platform that had one is vendor independence. An official cited the goal of not being locked to a single supplier. That is a standard procurement argument, and in this case it is also a hedge. A single model on a single platform is a single point of failure, and the department has learned that the companies behind these models change their minds about who they will serve.
That lesson is where Anthropic comes in. Claude, the company’s model, is not on the list. Anthropic and the department broke over disagreements about safety guardrails, and the split has not healed. The department’s two new picks come from OpenAI and xAI, two firms that have taken a more permissive line on military use. The result is a platform whose lineup reflects which companies were willing to sign the terms as much as which models performed best in testing.
The certification both new models cleared is a barrier worth understanding. Impact Level 5 covers systems that handle controlled unclassified information, the category just below classified material. Reaching that level requires an accreditation process, and only a small set of commercial AI systems have gone through it. A model that lacks the certification cannot be used for most of the work the department does, which is why the list of approved vendors is short and why each addition is consequential.
For OpenAI, the entry extends a shift that began with loosening its rules on military work. The company that once barred its technology from weapons applications now supplies a model to the department’s own platform. For xAI, the placement puts Grok into an institution with a user base that dwarfs its consumer footprint. Both companies gain a reference customer whose decisions shape what the rest of government buys.
The choice also says something about how the federal government is buying AI. Rather than building a single model in house, the department is assembling a catalog and letting its users pick. That approach spreads the risk of any one model failing or any one vendor withdrawing, and it turns the platform into a competition that the labs have to keep winning. A vendor that falls behind on certifications or on willingness to serve can be dropped without rebuilding the system around it.
For Google, the change marks the end of a brief monopoly. Gemini’s military version was the sole occupant of the platform, and it now shares the space. Google still has the advantage of incumbency and the deepest integration, but the department has signaled that it will not depend on any one lab for the models its workforce uses daily.
The absent vendor is the clearest signal of all. Anthropic built a reputation on caution, and the department’s break with it over guardrails was reported as a dispute about how tightly a model should refuse certain requests. Whatever the specific terms, the outcome is that a major American lab is shut out of the largest enterprise deployment in the federal government, over a disagreement about safety, while two rivals take its place.
That outcome cuts both ways. Anthropic keeps its position as the lab most willing to say no, which matters to the customers who value that stance and to the regulators writing rules for the sector. It loses a deployment that would have validated its security work at scale and paid for it. The department, meanwhile, gets vendors more willing to accommodate its needs, and takes on the question of what those vendors will do when a request tests their own limits.
The next step is whether the arrangement holds. A platform with three approved vendors can add a fourth, and Anthropic’s absence is a business decision as much as a policy one. If the guardrail dispute is resolved, Claude could return. If it is not, the department’s platform becomes a map of which labs are willing to work under its terms, and the map will be read closely by every agency that follows.
For now, the announcement is a procurement notice with a political shape. The military’s AI platform now runs on models from three companies, one of which was already there, two of which just arrived, and one prominent lab is missing by choice and by consequence. The department has said it does not want to be locked to a single supplier. It has not said how many suppliers it needs, or which disagreements disqualify one.
