Pentagon adds ChatGPT Mil and Grok to GenAI.mil, with Anthropic absent

The Pentagon added OpenAI's ChatGPT Mil and xAI's Grok for Government to its GenAI.mil portal on August 31, 2026. Both cleared Impact Level 5, the accreditation that allows sensitive but unclassified government data. The most telling detail is which company is not on the list: Anthropic, whose Claude models a federal judge said were blocked illegally.

GenAI.mil is the department's own front door to commercial frontier models. DefenseScoop reports it launched in December 2025 so staff could use these models without pushing sensitive data through consumer services.

The user numbers mean different things

Three figures circulate, and they are not measuring the same thing.

Figure	What it counts	Source
3 million	Military members, civilian employees and contractors eligible under the planned rollout	TechCrunch, DefenseScoop, NOTUS
1.7 million	Unique users onboarded since the December 2025 launch	TechCrunch, NOTUS
1.5 million	Active users as of June 2026	DefenseScoop

So the platform reaches a fraction of the people entitled to it, and that fraction is still large. Very few enterprise AI deployments anywhere are onboarding users at that scale.

What cleared, and what it can hold

Impact Level 5 is the part worth knowing, because it sets what data a model may touch.

DefenseScoop reports both models are authorized to store and process non-public, sensitive unclassified data. IL5 sits inside the department's cloud rules and is stricter than the basic commercial sign-off. Reports on the launch list a FedRAMP High base, physical separation between tenants, and a rule that only U.S. persons may staff the service.

Access is narrow by design. The platform runs on the military's unclassified network, reached with a Common Access Card from government-furnished equipment. There is no consumer login.

DefenseScoop also names the model behind ChatGPT Mil as GPT-5.6 Terra. Grok for Government runs xAI's Grok frontier model.

The intended work is unglamorous. The reports list planning, policy, logistics, administration, travel forms, contracting opportunities, research, analysis and drafting. Grok is additionally described as covering market research analysis, supply-chain management and mission execution.

Google got there first. DefenseScoop reports Gemini for Government launched on the platform in February 2026.

The company that is missing

Anthropic's absence is not a commercial decision.

The Pentagon designated Anthropic a national security supply-chain risk and ordered Claude removed. DefenseScoop reports the designation came about six months before this launch, after Anthropic refused to remove safety guardrails that keep Claude away from autonomous weapons and domestic mass surveillance.

That designation did not survive review. DefenseScoop reports a federal judge ruled the Pentagon's actions "illegal and baseless" in August 2026, and NOTUS reports the court found the government had unlawfully retaliated. Anthropic sued and is still contesting the matter.

So the portal now offers three vendors, and the fourth is absent because it declined to weaken its own safety policy. A court has said the exclusion was unlawful, and the exclusion continues.

One naming note. DefenseScoop refers to the Department of War, and explains the department was rebranded from the Defense Department by executive order while a formal name change still requires Congress.

What this means for developers

If you sell software to government, the sign-off is the product. Two frontier models reached three million possible users because they cleared IL5, not because they won a benchmark. If a federal rollout is on your roadmap, treat FedRAMP and impact-level work as core engineering, not paperwork you bolt on later.

There is a design lesson in what these tools are actually for. The listed uses are document-heavy routine work: forms, drafting, logistics, research. That is where big rollouts are finding value. It is a useful correction if you are building agent features for business customers, because the dull workflows are the ones that get bought.

The Anthropic situation is worth watching for a reason beyond the politics. A vendor kept a safety restriction, lost a large customer, won in court, and still does not have the customer back. If you are picking a model provider for regulated work, that order of events shows how long a fight can outlast the ruling meant to end it. Ask what happens to your integration if your provider lands in that spot.

Finally, note that only three million people are eligible while 1.7 million have onboarded. If you build internal AI tools, that ratio is a more realistic benchmark for adoption than any vendor case study.