The Pentagon added ChatGPT and Grok to its own AI portal for 3 million staff

The Pentagon now runs its own ChatGPT. On August 31 the Department of War added OpenAI’s ChatGPT Mil and Starshield AI’s Grok for Government to GenAI.mil, the enterprise portal it opened in December with Google’s Gemini. All three sit on unclassified networks.

More than 1.7 million of the department’s 3 million-plus personnel have logged into GenAI.mil in the nine months since launch, according to the release. That makes it one of the largest single frontier-model deployments anywhere, government or private. The models sit behind a front end the department controls, so a logistics officer drafting a movement order isn’t pasting it into a consumer chatbot with a consumer privacy policy. Which vendors get through that door, and on what terms, has become the actual fight.

What’s running behind the login

Both new models are accredited for CUI at IL5, as far up the sensitivity ladder as a system can go without crossing onto a classified network. ChatGPT Mil serves GPT-5.4 Terra today with GPT-5.6 Terra queued behind it, and it ships the same chat, files, projects and custom GPT surfaces commercial users already know, DefenseScoop reported. Grok brings deep-thinking inference, three reasoning modes (Auto, Fast and Expert) and reusable “playbooks” pitched at acquisition research and supply-chain work.

The confirmed details, each from the department or the vendor:

• GenAI.mil launched in December 2025 with Google’s Gemini for Government and has onboarded more than 1.7 million unique users since.
• Both new models cleared CUI accreditation at IL5 and are scoped to unclassified work: planning, policy, logistics and administration.
• OpenAI says data processed on GenAI.mil stays isolated in the government environment and never trains its public or commercial models.
• Grok arrives via Starshield AI, the government unit of the merged SpaceX and xAI, which closed in February at roughly $1.25 trillion.
• The Army’s XVIII Airborne Corps drafted an exercise operations order in six weeks on the platform, against a normal six-to-nine-month cycle, per Military Times.

On the record, the department stayed abstract. “Access to a diverse suite of AI capabilities from across the resilient American technology stack will give warfighters the tools they need to act with confidence and safeguard the nation against any threat,” said an official who would only be identified as a War Department official, in the DefenseScoop account. OpenAI’s government lead Joe Larson was more specific, calling the rollout “grounded in our national security principles, pairing access to powerful tools with clear safeguards, meaningful human judgment, and accountability for high-stakes decisions.”

The parts nobody has published

Start with money. Neither release names a contract value for these deployments. The public figure is still the Chief Digital and Artificial Intelligence Office award round from July 2025, when the CDAO handed OpenAI, Anthropic, Google and xAI ceilings of $200 million each under prototype agreements, a combined $800 million if every option gets exercised. Whether GenAI.mil draws on those vehicles or separate ones hasn’t been stated.

There’s also no published evaluation data. The department describes accreditation and enterprise controls, and it has not released red-team results, hallucination rates, or a threshold a model has to clear before a million-plus people can query it. The guardrails on the vendor side are contractual rather than technical: OpenAI keeps control of its own safety systems and delivers through the cloud instead of wiring into weapons platforms, NOTUS reported.

Then there’s the empty chair. Claude isn’t on GenAI.mil. The Pentagon had labelled Anthropic a national-security supply-chain risk after the company refused to lift its limits on mass domestic surveillance and fully autonomous weapons. On August 27, U.S. District Judge Rita Lin ruled that designation unlawful and found the government had retaliated against Anthropic for publicly defending those limits. Four days later, the portal added two competitors and not Claude.

What this means for you

If you’re building an internal AI portal, GenAI.mil is now the biggest working proof that the multi-vendor pattern holds at scale: one front end, several models behind it, per-model accreditation, and a written term keeping prompts out of the vendor’s training set. That last clause is the one enterprise buyers keep failing to get on paper, and the department got it from OpenAI in public. For anyone selling into defense, IL5 is the gate before a single user sees the product, and it’s slow. OpenAI’s first CDAO award landed in June 2025; ChatGPT Mil went live 14 months later.

Watch whether Claude turns up on the portal now that the risk label is gone. That’s the test of whether the ruling changed procurement or just the paperwork.