The Pentagon runs its own ChatGPT and Grok. What that says
Inside GenAI.mil: ChatGPT Mil, Grok for Government, Gemini, and the Claude ban a judge ruled illegal. Five steps to check an AI firm's government work.
On 31 August 2026 the US Department of War, which most people still call the Pentagon, put two more chatbots in front of more than three million people. One is a custom ChatGPT. The other is Grok. Both sit on an internal portal, GenAI.mil.
If you pay for an AI subscription, your question here is narrow: whether the company you pay every month does work you would rather it did not, and whether that work reaches the app you type into. The first part is answerable from public documents. The second only as far as those documents go, so it pays to be strict about where they stop.
What the Department announced, and when
GenAI.mil opened with a single model. Google Cloud's Gemini for Government was named the first enterprise AI on the platform in an announcement dated 9 December 2025, and the Department's own release calls it "the first of several frontier AI capabilities to be housed on GenAI.mil".
Two more landed on 31 August 2026, each with its own release on war.gov. OpenAI's ChatGPT Mil is presented as the adoption phase of an enterprise partnership established in 2025, built around "chat, files, projects, and custom GPTs" and aimed at document-heavy unclassified work: planning, policy, logistics, administration. Starshield AI's Grok for Government arrived in different vocabulary. The Department said it "will provide the warfighter with immediate productivity gains, stronger knowledge continuity and more secure and efficient collaboration", and that it "enables the Joint Force to execute missions faster and with greater precision". Starshield AI is SpaceX's secure satellite network.
Everything on the portal is accredited for Controlled Unclassified Information at Impact Level 5, an unclassified tier. The Grok release adds: "Since its launch nine months ago, GenAI.mil has already onboarded over 1.7 million unique users out of the Department's more than 3 million personnel." TechCrunch's report of 31 August 2026 cites the same release for it.
Careful. Lists of "AI companies working with the Pentagon" tend to blend two things. Three models are on GenAI.mil. Amazon Web Services, Microsoft, Nvidia and Reflection AI have Pentagon AI agreements of their own, reported by TechCrunch as deals to deploy AI on classified networks. Those are separate arrangements, and a company in the second group is not thereby on the portal.
The company that is not on the list
Claude is absent, and the reason is documented at length. In late February 2026 the Department pressed Anthropic to drop restrictions on how Claude could be used, and Anthropic declined. On 27 February and 3 March 2026, according to the court record, the President and the Secretary of War designated Anthropic a supply chain risk to national security, ordered federal agencies to stop using its products, and barred defence contractors from doing business with it, even on work unrelated to the military. A letter confirming the designation reached Anthropic on 4 March 2026. In a statement dated 5 March, chief executive Dario Amodei wrote that "Our only concerns have been our exceptions on fully autonomous weapons and mass domestic surveillance", and said the company was "very proud of the work we have done together with the Department".
A federal judge then voided much of that. In an order dated 27 August 2026, reported on 28 August, US District Judge Rita Lin, Northern District of California, granted Anthropic summary judgment on its First Amendment, due process and Administrative Procedure Act claims. It calls the measures imposed on Anthropic "illegal and baseless", and describes the government's central theory, that Anthropic could alter or disable models already deployed in Pentagon systems, as "entirely unfounded": those models are static and beyond the company's reach, which the government did not dispute. "An IT vendor does not become a potential adversary of the United States whenever it asks probing questions," Lin wrote, and: "The empty invocation of national security is not a blank check to punish and retaliate against government critics."
Two details stop this being a clean ending. Anthropic brought two challenges over two separate designations, and Engadget reports that it stays designated a supply chain risk pending the second. The order also stresses that the Department remains free to stop buying Anthropic's technology through ordinary procurement. Anthropic's response, given to CNBC and quoted by Engadget: "We welcome the court's ruling that this supply chain risk designation was unlawful."
Whether any of this reaches the app you use
Each deployment is described by its operator as a walled-off environment. OpenAI's own page says "Data processed on GenAI.mil remains isolated to the government environment and is not used to train or improve OpenAI's public or commercial models". Google Cloud's announcement says "The data used by the DoW on this platform is never used to train Google's public models."
Both sentences run in one direction: government data stays out of commercial models. Nothing on any page fetched for this piece addresses traffic the other way. So, the honest summary. These are separate deployments in authorised government cloud infrastructure with their own accreditation and terms, and no page we read describes your consumer conversations as an input to them. We will not fill the remaining silence with a guess.
Warning. A vendor's page is the vendor's account of itself, written about the government deal. Where your own chats go is governed by your consumer terms, a separate document with separate wording. Never read a reassurance about a military contract as a statement about your account.
Where we stand
We are a Belgian consumer subscription with no government or defence product. We route to OpenAI, Google, Anthropic, xAI and Mistral AI directly, and to open-weight models through OpenRouter, all named on our live privacy policy. We do not train on your conversations. What the provider behind a given model does with them varies by model, and the models where we know of a difference carry a notice on their info card in the picker.
Five steps to check a company's government work yourself
Work down these in order and you will know more than most coverage does.
Start at the company's own government page. OpenAI publishes its GenAI.mil deployment under OpenAI for Government, Anthropic posts its side of the dispute as dated newsroom statements, and Google Cloud runs a public sector press page. All advocacy, and still the fastest route to which product, agency and date.
Read the buyer's release alongside the seller's. The Department publishes its announcements at war.gov, and they frequently say more than the vendor did. Both 31 August releases named capabilities and accreditation levels, and the Grok one supplied the 1.7 million figure most coverage then quoted. Mind the naming: the buyer's releases say Department of War, most reporting still says Pentagon or Department of Defense.
Check the procurement record. Awards, amounts and dates for US federal contracts are public at USAspending.gov, and solicitations at SAM.gov. A company can decline to discuss a contract, but it cannot unpublish one. This step separates a signed deal from an announced intention.
When there is a dispute, read the docket. The 59 pages behind a "judge slams Pentagon" headline are free. CourtListener serves federal filings as PDFs, and the Anthropic order is there in full. Reading it produced facts no summary carried, including the government conceding the point its risk assessment rested on.
Read a denial for its scope rather than its tone. Take OpenAI's sentence above. Its verb is "train or improve", its object is "public or commercial models", and it covers data processed on GenAI.mil. Narrow, specific and checkable, which is what makes it worth something. Compare "we take your privacy seriously", which carries no noun, no verb and no product. Where a statement declines to name the product, the data and the direction, treat it as unanswered rather than as a no.
All four companies named here, the three on the portal and the one kept off it, also serve ordinary customers. The linked documents set out the arrangements if you want to read them before deciding how you feel about them.