ChatGPT Mil and Grok on GenAI.mil: IL5, GPT-5.4 Terra

Nine months in, GenAI.mil ran on one model. Google's. Since Monday it runs on three: the Department of War switched on OpenAI's ChatGPT Mil and Starshield AI's Grok for Government on 31 August, both cleared at Impact Level 5. Reads like the week the Pentagon went all in on frontier AI. Mostly it isn't. IL5 tops out at sensitive unclassified data, the Controlled Unclassified Information tier, which is the paperwork half of the building rather than the interesting half. And the ChatGPT that landed is GPT-5.4 Terra, one version behind what OpenAI sells to the rest of us.

Three models now, and the tier they all share

GenAI.mil went live on 9 December 2025 with Google’s Gemini for Government and nothing else. Monday’s addition brings OpenAI’s ChatGPT Mil and Starshield AI’s Grok for Government onto the same portal, which means the Department’s Chief Digital and Artificial Intelligence Office has now lit up three of the four frontier vendors it holds contracts with.

The detail worth pinning down is IL5. It’s a Defense Department cloud accreditation level, and it caps what you may put in the box: non-public sensitive unclassified information, Controlled Unclassified Information included. Classified material lives higher up the stack and is not in scope here. So when the coverage says the military has ChatGPT now, what it has is an accredited environment for the document work. Planning packets. Policy drafts. Acquisition market research.

That’s not a knock. Honestly, CUI-tier drudgery is where a chat model earns its keep, and it’s the part of a 3 million person organisation that eats the most hours. But if you read the announcement expecting the interesting half of the building, it isn’t there.

What each one ships

ChatGPT Mil gives you chat, files, projects and custom GPTs, plus an Offline Search capability that comes back with citations. The model behind it is GPT-5.4 Terra, and DefenseScoop reports GPT-5.6 Terra as rolling out. One version behind at launch.

Grok for Government leads with reasoning instead: deep-thinking inference with Auto, Fast and Expert modes, customisable workspaces, persistent projects, plus reusable playbooks for repeated workflows.

Neither list is exotic, and that’s the point of an accredited build. Accreditation is granted against a specific configuration, so every model swap and every feature flag is a paperwork event, which is exactly why the government edition trails the commercial one and probably always will. Worth remembering the next time somebody promises you a same-day model upgrade inside a regulated environment.

The Claude-shaped hole, and the clause everyone forgot

Here’s the bit that made me reread the wire copy.

Anthropic is one of the four frontier vendors CDAO signed. It’s not on the platform, because the Department designated it a supply chain risk in early 2026 and pushed it out of federal eligibility. On 27 August a federal judge vacated that designation and stacked a permanent injunction on top. Four days later, this announcement, without Claude in it. Which tracks with what the order actually does: restoring a vendor’s eligibility isn’t the same as awarding that vendor anything.

The underlying fight was about usage restrictions. The Department wanted its models available for all lawful purposes, Anthropic kept carve-outs on lethal autonomous weapons and mass domestic surveillance, and the ban followed.

Now go read OpenAI’s published national security principles. They commit, contractually, to no use of the technology for mass domestic surveillance, no use to direct autonomous weapons systems, and no high-stakes automated decisions without meaningful human judgment. Joe Larson, OpenAI’s VP and head of government, put Monday’s integration squarely on those principles.

I might be reading too much into two documents I only have the public version of. But from outside, those restrictions look close enough that the difference in outcome is hard to explain by the restrictions alone. Something else decided it, and the public record doesn’t say what.

Is anyone actually using it

More than I expected, which is the genuinely interesting number here.

1.7 million unique users onboarded out of roughly 3 million eligible, up from 1.5 million reported in June. Enterprise software rarely posts adoption like that, and part of the reason is that it wasn’t optional everywhere: a January memorandum made GenAI.mil the mandated path for the Navy and Marine Corps, with a transition deadline of 30 April.

The concrete use case doing the rounds comes from the Army’s XVIII Airborne Corps, which produced a full exercise operations order for U.S. Southern Command in six weeks. The Department’s framing is that the traditional path takes six to nine months and a bigger staff. Take that comparison as the Department’s, not as an independent measurement, because nobody has published the methodology behind it.

If you build for public sector customers, the transferable lesson is duller than the headline and more useful. The accreditation tier decides your feature set, the mandate decides your adoption curve, and the model version you ship is whatever survived the paperwork. None of that changes because the vendor logo does.

We’ll come back to this when GPT-5.6 Terra is confirmed live rather than rolling out, or when a fourth model appears. For what the same GPT-5.6 family is doing on the commercial side, see our notes on ChatGPT’s Site tools and WebMCP.

Sources

The launch, the IL5 accreditation and the user numbers come from DefenseScoop and TechCrunch, both dated 31 August 2026. The XVIII Airborne Corps example, the Navy mandate and the data isolation statement are from Military Times. The original platform launch on Gemini for Government is documented in DefenseScoop’s December 2025 report. OpenAI’s restrictions are published in Our approach to government and national security partnerships. Our earlier piece on the court order is Anthropic v. Pentagon: supply chain risk label vacated.

Frequently asked questions