GenAI.mil is the US Department of Defense's central generative AI portal for its roughly 3 million military, civilian and contractor personnel. It launched about nine months before August 2026 with Google's Gemini as the first available model, and had onboarded more than 1.7 million unique users by the time OpenAI's ChatGPT Mil and Starshield AI's Grok for Government were added on August 31, 2026. It is scoped to controlled unclassified information, not classified work.

What does IL5 mean for AI models?

Impact Level 5 is the Department of Defense's authorisation tier for controlled unclassified information, including unclassified National Security System data — the highest level short of classified environments. Department press releases indicate all three models on GenAI.mil have cleared IL5. In practice it dictates where data is processed, who can access the infrastructure, and how the deployment is isolated, which is why these are separate government builds rather than the commercial products.

Can Pentagon staff use these for classified work?

No. GenAI.mil covers controlled unclassified information (CUI), not classified material. Personnel can draft, summarise, analyse and code against sensitive-but-unclassified data. Classified workflows run on separate accredited systems, and the IL5 authorisation these models hold is explicitly the tier below classified environments.

Why does the Pentagon run three competing models instead of one?

Multi-vendor sourcing avoids lock-in on infrastructure the department cannot easily migrate, and it lets each model be routed to the work it handles best. It also reflects procurement reality: no single vendor should hold a monopoly on a capability used by 3 million people. The practical effect is that model choice becomes a user decision inside one accredited portal rather than a department-wide bet.

Quick Answer

ChatGPT Mil vs Grok for Government vs Gemini: GenAI.mil

Published: September 2, 2026

The Short Answer

On August 31, 2026, the Pentagon deployed OpenAI’s ChatGPT Mil and Starshield AI’s Grok for Government across GenAI.mil, joining Google’s Gemini, which had been the portal’s first model since roughly late 2025.

The numbers that matter:

3 million DoD military, civilian and contractor personnel are eligible.
1.7 million unique users had already onboarded before the expansion.
All three models cleared IL5 — the DoD authorisation tier for controlled unclassified information.
Scope is CUI, not classified.

Last verified: September 2, 2026.

Side by Side

	ChatGPT Mil	Grok for Government	Gemini
Vendor	OpenAI	Starshield AI (SpaceX)	Google
Added to GenAI.mil	Aug 31, 2026	Aug 31, 2026	First — ~late 2025
Authorisation	IL5	IL5	IL5
Data scope	CUI	CUI	CUI
Classified use	No	No	No
Incumbency	New	New	~9 months of usage data

The three arrive on very different footing. Gemini has nine months of real usage across a workforce that has already produced 1.7 million accounts. ChatGPT Mil and Grok for Government start from zero inside an environment where habits are formed.

What IL5 and CUI Actually Constrain

These two acronyms do most of the work in this story, and most coverage skips them.

CUI — controlled unclassified information — is the category covering sensitive government data that is not classified: contract details, logistics, personnel administration, unclassified technical documentation, internal planning. It is the overwhelming majority of what a large organisation does day to day.

IL5 — Impact Level 5 — is the DoD authorisation tier covering CUI including unclassified National Security System data. It is the highest tier short of classified environments, and it imposes concrete requirements: where compute physically sits, who can administer it, how tenancy is isolated, and how data is prevented from flowing back to the commercial product.

That last point is the reason these are named differently from their consumer counterparts. ChatGPT Mil is not ChatGPT with a different logo — it is a separately accredited deployment whose isolation properties are the actual product. The model underneath may be closely related; the guarantees around it are not.

What it does not cover: classified work. Anyone reading this as “the military is running its war planning through ChatGPT” has misread the accreditation. Classified workflows run on separate accredited systems.

The Adoption Number Deserves Attention

1.7 million unique users out of roughly 3 million personnel is a 57% penetration rate in about nine months, on a single-model portal, inside an organisation not known for rapid software adoption.

For comparison, most enterprise AI rollouts in 2026 struggle to move past pilot groups. The DoD figure suggests that when a large organisation removes the two real blockers — is it approved, and is it already available — usage follows without persuasion. The bottleneck in enterprise AI adoption has rarely been enthusiasm; it has been procurement and accreditation.

One caveat on the metric: “onboarded unique users” counts accounts that have accessed the portal, which is a weaker signal than sustained weekly use. It measures removed friction, not embedded workflow.

The Grok Deployment Is Contested

Worth stating plainly, because it is part of the record: the Grok deployment drew criticism, including reporting that engineers had raised concerns about content-safety failure modes without a reliable fix before the rollout proceeded. Coverage of that dispute is journalistic reporting on internal deliberations, not a government finding.

The broader pattern is more consequential than any single vendor dispute. Frontier AI is moving into government infrastructure faster than the evaluation frameworks for it are settling. The same week the Pentagon expanded to three models, the EU designated ChatGPT a Very Large Online Search Engine under the DSA and OpenAI confirmed its forthcoming Astra model had reached “Critical” cyber capability. Deployment is outpacing governance in both directions at once.

What This Means Outside Defence

For enterprise buyers: IL5 accreditation for three frontier models is a strong signal for regulated industries generally. Finance, healthcare and critical infrastructure procurement teams asking “has anyone with real requirements accredited this” now have an answer.

For the multi-model question: the largest single AI deployment in the US government chose three vendors rather than one. If your organisation is debating standardising on a single model provider, the most security-constrained buyer available concluded that routing across several was the better architecture.

For vendors: government is now a primary distribution channel, not a side market. 3 million eligible seats in one portal is larger than most enterprise contracts, and incumbency inside an accredited environment is extremely durable.