ChatGPT Mil and Grok Take the Desk Anthropic Lost

The Department of War put ChatGPT Mil and Grok for Government on GenAI.mil on August 31, 2026, three days after a federal judge struck down its Anthropic blacklist. Both tools are cleared for Controlled Unclassified Information at Impact Level 5 and now share the department’s generative AI portal with Google’s Gemini.

The dual rollout is the adoption phase of a 2025 OpenAI partnership and a Starshield AI delivery of Grok. It also shows which labs now own the unclassified desk, and which lab still does not.

Three Models on GenAI.mil, Claude Still Out

Secretary of War Pete Hegseth introduced GenAI.mil to the workforce in December 2025 as a secure generative AI platform for every member of the department, with Gemini as the first model. A Pentagon news note on September 1 said the new OpenAI and Starshield tools give staff a first, second, and third opinion on the same portal.

ChatGPT Mil is the adoption phase of an enterprise partnership the department struck with OpenAI in 2025. The August 31 release says the core experience is chat, files, projects, and custom GPTs, with more features to follow, and that the tool is built for more than 3 million personnel doing document-heavy unclassified work in planning, policy, logistics, and administration.

Starshield AI’s Grok for Government arrived the same day, with deep-thinking inference, Auto, Fast, and Expert reasoning modes, custom workspaces, persistent projects, and reusable playbooks meant to hold group knowledge. The department pitched it at acquisition research and supply-chain work for logisticians.

GENAI.MIL AT THE DUAL LAUNCH

• Unique users: The Grok release counts over 1.7 million unique users since the portal opened.
• Workforce size: Both launches are scoped to more than 3 million military and civilian personnel.
• Clearance: ChatGPT Mil and Grok for Government are accredited for Controlled Unclassified Information at Impact Level 5.
• Age of the portal: The department said GenAI.mil had been live nine months when the two models arrived.

Those 1.7 million accounts were built mostly on Gemini. ChatGPT Mil and Grok inherited a crowded desk, not an empty one. The Chief Digital and AI Office, which runs the portal with the Office of the Under Secretary of War for Research and Engineering, said it will keep scaling frontier models across the Joint Force, a track that already includes a hiring push for AI engineers.

OpenAI’s Mohammed Husain, strategic delivery lead for cyber, had said in June that ChatGPT would go live in early July. The actual switch-on was August 31. Live and on-demand training for each model is free to department staff.

Anthropic’s Court Win Left Claude Off the Portal

Anthropic was in the same cohort a year earlier. On July 14, 2025, the company said the Chief Digital and AI Office had awarded it a prototype other-transaction deal with a $200 million ceiling over two years. A later court filing by the company said Google, OpenAI, and xAI received matching awards, and that Claude had been the first frontier model cleared for classified government work.

That path closed in late winter. Anthropic agreed to drop most use limits, then refused two: lethal autonomous warfare and mass surveillance of Americans. On February 27 and March 3, 2026, President Donald Trump and Hegseth moved to designate the company a supply chain risk, a procurement brand usually aimed at foreign sabotage, and to pull Claude out of federal workflows.

The empty invocation of national security is not a blank check to punish and retaliate against government critics.

Judge Rita F. Lin, U.S. District Court, Northern District of California

Lin ruled on August 28, 2026, that the designation was unlawful First Amendment retaliation and arbitrary. She wrote that the department is free to pick its AI vendor, and that the broad measures against Anthropic were illegal and baseless. Three days later, ChatGPT Mil and Grok for Government went live on the portal Anthropic’s own filing said it had lost the chance to join.

HOW THE VENDOR STACK SHIFTED

1. July 14, 2025: Anthropic announces the two-year CDAO award with a $200 million ceiling, in a round that also included Google, OpenAI, and xAI.
2. December 2025: GenAI.mil opens with Gemini for Government, cleared for Impact Level 5.
3. February 9, 2026: The department and OpenAI announce a ChatGPT integration after the portal passes one million unique users in two months.
4. February 27, 2026: Trump and Hegseth start the supply chain risk process against Anthropic.
5. March 3, 2026: Hegseth signs the designation under 10 U.S.C. § 3252.
6. May 1, 2026: Eight firms get deals to put models on Impact Level 6 and 7 classified networks. Anthropic is not on the list.
7. August 28, 2026: Lin strikes down the blacklist.
8. August 31, 2026: ChatGPT Mil and Grok for Government go live on GenAI.mil.

Commerce Secretary Howard Lutnick later said, “We trust Anthropic,” and that the company had done what the administration asked and was “back on the right side.” The portal still does not carry Claude. The Pentagon case still hanging over Anthropic is now a split inside the same administration, not a restored login.

Who Holds the Classified Network Seats?

On May 1, 2026, the department named eight companies for classified Impact Level 6 and 7 networks: SpaceX, OpenAI, Google, NVIDIA, Reflection, Microsoft, Amazon Web Services, and Oracle. The release said the deals were meant to make the United States military an “AI-first fighting force” and to avoid vendor lock. Anthropic was not named.

By then GenAI.mil had 1.3 million users, tens of millions of prompts, and hundreds of thousands of agents in five months. The unclassified portal and the classified stack are different rungs. The same labs that said yes to the department’s use terms show up on both.

WHO IS ON THE DESK AND WHO IS NOT

NVIDIA, Reflection, Microsoft, Amazon Web Services, and Oracle fill out the eight on classified networks and do not appear on this portal. Google kept first chair on GenAI.mil. OpenAI and Starshield AI took the next two seats. Claude is on neither board.

Chat, Files, Playbooks and the Desk Work

Officials have said they do not want the GenAI.mil models used for combat. The work the vendors describe is staff work: files, drafts, research, and memory that used to live in shared drives and inboxes.

OpenAI’s February 9, 2026 note said ChatGPT on the portal runs in authorized government cloud, with data isolated so it is not used to train public or commercial models.

WHAT CHATGPT MIL IS CLEARED TO HANDLE

• Policy files: Summarizing and analyzing policy and guidance documents.
• Contracts: Drafting and reviewing procurement and contracting materials.
• Compliance: Generating internal reports and checklists.
• Staff support: Research, planning, mission support, and administrative workflows.

Grok for Government is pitched as a different shape on the same desk. Auto, Fast, and Expert modes let a user trade speed for depth. Persistent projects and playbooks are meant to keep a unit’s method in the tool so the next shift does not start from zero. The department’s own examples are market research for buyers and supply-chain management for logisticians, not a fire-control loop.

That is the product the 3 million-person workforce is being trained to open. It is also the product that can become habit before Congress finishes a weapons rule.

OpenAI Already Agreed to All Lawful Uses

The fight that took Claude off the field was not about chat, files, or playbooks. It was about whether a lab could keep two hard noes after the department asked for all lawful military uses, including the two Anthropic would not give: lethal autonomous warfare and mass surveillance of Americans.

OpenAI’s GenAI.mil post describes model and platform safeguards and then says the system is supporting all lawful uses. That sentence is the bargain Anthropic would not sign in full. The unclassified portal does not, by its own billing, fire weapons. The vendor that wanted a seat on a 3 million-person desk still had to accept the department’s wider use terms to get there.

Lin’s order left the department free to choose vendors. It forbade using a national-security stamp to punish a lab for saying no in public. The August 31 launches are the choice the order allowed. They are also the map of who accepted the terms and who did not.

The louder objection around the rollout is not battery life or a missing button. It is that Claude is absent, and that the labs on the portal are the ones that did not keep Anthropic’s two red lines. Desk work is the visible layer. The contract language underneath is what decided who got a login.

Senate Bills on Nuclear Use, Surveillance and Autonomy

The two limits Anthropic tried to keep in a contract are the same limits some senators are trying to write into statute, after the fact, for the whole department.

Sen. Elissa Slotkin, D-Mich., introduced the AI Guardrails Act of 2026 (S.4113) on March 17, 2026. The bill would bar the department from using AI to launch or detonate a nuclear weapon and from using AI to monitor, track, profile, or target people in the United States without an individualized legal basis.

Sen. Kirsten Gillibrand, D-N.Y., followed on June 2, 2026, with the Secure and Accountable Military AI Act. Her release describes an absolute prohibition on nuclear target selection and on executing a nuclear launch, a ban on person-based tracking of U.S. persons at home, with narrow cyber and force-protection carve-outs, and a general bar on fully autonomous weapons except for some supervised local-defense systems. High-consequence tools would need written approval from an under secretary or the vice chairman of the Joint Chiefs.

LIMITS THE TWO BILLS WOULD WRITE

• Nuclear use: Both bills try to keep AI off launch and, in Gillibrand’s text, off target selection for nuclear weapons.
• Domestic tracking: Slotkin and Gillibrand each aim to block AI-driven monitoring of people in the United States without a legal basis.
• Autonomy: Gillibrand would generally bar fully autonomous weapons; Slotkin’s bill is built around human judgment on lethal force.
• Sign-off: Gillibrand would require senior written approval before high-consequence AI goes operational, including lethal targeting support and cyber.

Gillibrand said the department is moving toward powerful AI “without commonsense guardrails in place.” The bills are not law. The portal is. The labs that refused to bake those limits into a vendor contract are watching from outside the login screen while the labs that did not refuse are already on it.

The Supply Chain Label After the Ruling

On September 3, 2026, Under Secretary of War for Research and Engineering Emil Michael posted that the designation was still in force for the department and the defense industrial base, using the same “thank you for your attention to this matter” close associated with Trump.

Anthropic is still a designated Supply Chain Risk at @DeptofWar and for the Defense Industrial Base. Thank you for your attention to this matter!

— Under Secretary of War Emil Michael (@USWREMichael) September 3, 2026

Replies pointed out that Lin had already ruled and that the department appeared to be ranking labs by who accepted looser use terms. Lutnick’s trust line and Michael’s label cannot both describe the same working relationship. Only one of those positions matches the portal as built: Gemini, ChatGPT Mil, and Grok for Government are on GenAI.mil, and Claude is not.

The court told the department it could pick its vendors. On August 31 it did. The research chief is still using the stamp the court said it could not use. That is the desk OpenAI and Starshield AI now share, and the one Anthropic still does not.