The Pentagon Now Runs ChatGPT and Grok Side by Side for Three Million People

On Monday, the Pentagon quietly became the largest enterprise deployment of frontier AI in the world. With two press releases issued August 31, the Department of Defense — rebranded by executive order as the Department of War — switched on OpenAI's ChatGPT Mil and Starshield AI's Grok for Government inside GenAI.mil, the department's centralized generative AI portal. Roughly 3 million military members, civilian employees and contractors now have three frontier models sitting behind a single login, and can toggle between them the way a knowledge worker toggles browser tabs.

The scale is the story. GenAI.mil launched on December 9, 2025 with Google's Gemini for Government as its only model. Nine months later the department says more than 1.7 million unique users have onboarded — better than half the eligible workforce, a rate of enterprise AI adoption that few Fortune 500 companies have matched. Both new models are accredited to handle Controlled Unclassified Information at Impact Level 5, the highest authorization DoD grants for non-public, sensitive unclassified data. OpenAI and Elon Musk's SpaceXAI each hold defense contracts worth up to $200 million for the work; Google and Anthropic received awards in the same tranche.

The two tools were sold in strikingly different registers. ChatGPT Mil, an outgrowth of the OpenAI for Government program, is pitched as an office productivity layer. Its core experience "centers on chat, files, projects, and custom GPTs, with additional features sequenced over time," Pentagon officials wrote, with the tool aimed at document-heavy unclassified work in planning, policy, logistics and administration — procurement drafting, travel forms, responses to contracting opportunities, compliance checklists. A person familiar with the work at OpenAI told DefenseScoop that personnel currently reach GPT-5.4 Terra and will "soon be able to chat directly with GPT-5.6 Terra," alongside an offline search capability that returns detailed citations. Data stays inside the environment and is not used to train the models.

"ChatGPT on GenAI.mil will make unclassified work throughout the Department more efficient, freeing up time and resources so DoW personnel can focus on the mission," said Joe Larson, OpenAI's vice president and head of government and a former senior Pentagon official, in a statement to Fortune. In a separate statement he framed the deal in safety terms: "This integration is grounded in our national security principles, pairing access to powerful tools with clear safeguards, meaningful human judgment, and accountability for high-stakes decisions."

Grok got the warfighting language. The department's release promises "deep-thinking inference, adaptive reasoning modes (Auto, Fast, and Expert), customizable workspaces, persistent projects and reusable 'playbooks' that capture and scale institutional knowledge," enabling the Joint Force "to execute missions faster and with greater precision" across contexts from acquisition market research to supply-chain management. Spokespeople for xAI and SpaceX did not respond to DefenseScoop's requests for comment.

Why this matters

Start with who is not on the platform. Anthropic's Claude was originally slated for GenAI.mil and is conspicuously absent, the casualty of a public rupture over whether a vendor may restrict military use of its own product. Anthropic sought contractual guardrails against mass domestic surveillance and fully autonomous weapons; the Pentagon demanded unrestricted access for any purpose its leaders deem lawful, designated the company a supply-chain risk, and ordered its models purged from national security systems. Last week a federal district judge ruled that designation "illegal and baseless." OpenAI, by contrast, negotiated a separate agreement that preserves limits on surveillance, autonomous weapons and certain high-stakes decisions without a human in the loop — enforced, the company says, through contractual and technical means, including keeping its safety systems under its own control and serving models from the cloud rather than on weapons.

That is the procurement precedent worth watching: a vendor that negotiated quietly kept its guardrails and its contract; a vendor that negotiated publicly got blacklisted and had to win in court. NOTUS also noted a gap the announcements do not close — Monday's releases say ChatGPT Mil supports "planning" without specifying whether that includes operational planning, intelligence analysis or targeting. Asked directly, a Pentagon official would only reiterate the IL5 CUI accreditation.

Then there is the Grok question, which requires care. In September 2025, Business Insider reported that twelve current and former xAI workers said they regularly encountered sexually explicit material, including AI-generated child sexual abuse material, in data-annotation work on Grok. Senator Elizabeth Warren cited that record in a March 16, 2026 letter to Secretary Pete Hegseth about granting Grok access to classified systems: "I am concerned that Grok's apparent lack of adequate guardrails could pose serious risks to the safety of U.S. military personnel and to the cybersecurity of classified systems." Her letter also asserts that the NSA "conducted a classified review" and found Grok posed concerns other models did not.

The sharper claim now circulating — that DoD deployed Grok on GenAI.mil after xAI engineers concluded internally that CSAM generation has no reliable technical fix — traces to a single low-credibility outlet, Tech Times, and The Vault found no corroboration for that specific sequencing in DefenseScoop, TechCrunch, Fortune, NOTUS or the department's own releases. Treat the underlying safety record as documented and the causal framing as unverified.

Data governance is the quieter risk. In June, the Defense Counterintelligence and Security Agency warned that unauthorized "shadow AI" could create leaks; Congress has separately ordered the Pentagon to assess cyber risk from both authorized and unauthorized AI software. Hegseth's December memo left little room for opting out: "I expect every member of the Department to log in, learn it, and incorporate it into your workflows immediately."

Watch three things next. Whether Claude returns to GenAI.mil now that the supply-chain designation has been struck down. Whether the department publishes usage rules distinguishing administrative drafting from planning, analysis and targeting. And whether Warren's oversight questions — still unanswered publicly — produce the xAI safety documentation Congress has asked for twice.