Pentagon clears ChatGPT and Grok onto GenAI.mil alongside Gemini

The person this lands on is whoever has to turn a stack of unclassified movement paperwork into something a colonel will read by Friday, and who now sees more than one approved place to put it ^[2][12]. The department's stated reason for the expansion is that each top-tier provider has distinct technical strengths worth matching to specialized missions ^[3]. The pitch for ChatGPT Mil in the department's own release is narrower and more candid: a "familiar commercial experience" inside the secure environment, built around chat, files, projects and custom GPTs, with other features sequenced later ^[12][13].

That gap is the whole story of multi-model buying. Teams assume users compare model strengths per task, but users mostly open the interface their thumbs already know, which is why familiarity was worth putting in a press release at all ^[12]. Both models cleared Impact Level 5, the highest authorization for non-public, sensitive unclassified data, so the boundary is the same and the difference the user experiences is the front door ^[10].

The roadmap's arithmetic is less flattering than the pace it was announced with ^[11]. December's plan named four models for a platform meant to reach more than three million people ^[5][6]. Three are live roughly eight months later ^[16][17], and the fourth is absent for reasons unrelated to capability: Anthropic pressed for contractual guardrails against mass surveillance of Americans and lethal autonomous weapons, while the Pentagon wanted unrestricted access for what it deems lawful purposes ^[1]. The department then designated Anthropic a supply-chain risk and ordered its products purged from national security systems, and last week a federal district judge called that illegal and baseless, per DefenseScoop ^[7][8]. The same official who promises an architecture that prevents AI vendor lock is running a portal where one named vendor was removed by designation rather than by evaluation ^[9].

For anyone about to defend a second assistant licence internally, the useful sort is two questions crossed. First: is the second model there for a task the first genuinely cannot complete, or for a task the first can complete but people do not trust it to? Second: who picks, the person at the keyboard or an administrator default? Capability paired with an admin default gives you a routing decision you can measure. Habit paired with user choice is a real purchase too, just one you should call habit rather than dress up as coverage, while capability paired with user choice quietly turns every employee into a model-selection expert, and habit paired with an admin default just means paying twice for the same output.

The readings that settle it are repeat use by the same person on the same task the following week, time to a first draft someone actually sends, and the share of documents that leave the tool without being redone by hand. Seats provisioned and sessions opened will look excellent on all three products and tell you nothing. If the honest answer for your second model is habit, buy it and say so; save the capability argument for the one that earns it.