On August 31, the Department of War confirmed that military-tailored versions of OpenAI’s ChatGPT and xAI’s Grok are now live on GenAI.mil, the Pentagon’s internal generative AI portal. Both join Google’s Gemini, which has anchored the platform since its December 2025 launch. Of the roughly 3 million civilian, military, and contractor personnel eligible for access, the department says about 1.5 to 1.7 million have already logged in — a genuinely large deployment for enterprise AI, let alone one running inside a defense bureaucracy.
The headline framing is “the military gets ChatGPT and Grok.” The more interesting story is who is conspicuously absent from that sentence: Anthropic. Claude was part of the same 2025 prototype cohort that produced GenAI.mil in the first place, alongside Gemini, ChatGPT, and Grok. It never made it into this expansion, and the reason is not a technical one — it is a fight over what the model is allowed to be used for.
A Contract Won by Dropping Guardrails
According to reporting on the rollout, Anthropic had insisted on contractual language restricting its models from use in mass surveillance or lethal autonomous weapons systems before it would grant the Pentagon broader access to Claude. The Pentagon did not accept those terms. Roughly six months before this expansion, the department had gone further and formally labeled Anthropic a “supply-chain risk” — a designation a federal judge has since ruled “illegal and baseless,” according to court findings referenced in coverage of the dispute. OpenAI and xAI, whose products now anchor the platform’s newest tier, did not attach the same conditions to their government contracts.
That is the actual trade being made here, and it is worth stating plainly rather than letting it sit as a footnote: a defense customer with enormous purchasing power effectively selected its AI vendors based on which labs were willing to remove use-case restrictions, not which labs offered the strongest technical fit. Anthropic held a stated position — no support for autonomous lethal targeting, no mass surveillance use — and it cost the company a contract serving 3 million users. OpenAI and xAI did not hold that same line, and they got the business.
What’s Actually Live
ChatGPT Mil ships alongside GPT-5.6 Terra, described as the current frontier tier, running next to the previously deployed 5.4 Terra version so personnel can choose between them. It includes chat, file handling, persistent projects, custom GPTs, and an “Offline Search” capability built around detailed source citations — a meaningful detail for a bureaucracy that runs on paperwork and needs to trace where an AI-generated answer came from. Grok for Government, delivered through xAI’s Starshield AI arm, ships with deep-thinking inference and three adaptive reasoning modes — Auto, Fast, and Expert — plus customizable workspaces and “playbooks,” reusable prompt templates meant to capture institutional knowledge that would otherwise walk out the door with a retiring officer.
Both products cleared Impact Level 5 authorization, the Pentagon’s classification threshold for handling sensitive but unclassified data — a real bar to clear, not a marketing claim, since IL5 governs what a system is legally permitted to touch. The stated use cases are unglamorous and exactly what you’d expect from an organization this size: drafting travel forms, generating contracting responses, supply-chain planning, market research for acquisition officers, and the general grind of document-heavy administrative work that eats a bureaucracy’s time.
The Pattern to Watch
This is the second time in 2026 that a frontier lab’s safety commitments have collided directly with a government’s appetite for unrestricted access, and the second time the lab that held firm lost the deal. It is a useful data point for anyone trying to model how AI vendor selection actually works once you move past consumer subscriptions and into institutional procurement: the safety terms a lab is willing to publicly commit to are not just a branding choice, they are a competitive variable that a government buyer will price in — and sometimes select against.
What This Means for Philippine Founders
This story matters less as “which chatbot the Pentagon uses” and more as a template. The Philippines is early in its own conversation about government AI adoption — the DICT and individual agencies are piloting generative tools for casework, and any founder building AI products aimed at government or large-enterprise clients here should watch this pattern closely: a buyer with enough leverage will ask you to strip out the exact restrictions that make your product defensible on safety grounds, and the vendor willing to say yes usually wins the contract, at least in the short run. If you’re a Filipino AI startup pursuing a government or BPO-scale enterprise deal, decide now, on paper, which use-case restrictions you will not remove for any contract size — because that question will eventually get asked, and having answered it in advance is very different from negotiating it live under deal pressure.
There’s a second, more direct angle too: GenAI.mil’s IL5 security bar and its “Offline Search with citations” feature are both a reasonable minimum spec for any AI tool being pitched into regulated Philippine sectors — banking, insurance, government casework — where BSP and SEC compliance teams will eventually ask the same kind of provenance question the Pentagon just answered for itself.
Share this article