Pentagon signs classified AI deals with OpenAI and Google — Anthropic excluded

The Pentagon just widened its AI vendor list for classified work — and the missing name is the whole story. On May 1, the Defense Department said OpenAI, Google, Microsoft, Amazon Web Services, Nvidia, SpaceX, and Reflection AI now have agreements to bring AI into its classified network environments. That is not a generic “AI for government” move. It is a push into the systems used for secret and top-secret defense work, where the stakes are operational and political at the same time. (cybernews.com) ### What actually changed? The new piece is scope. OpenAI had already announced a Pentagon agreement on February 28 for deploying advanced AI systems in classified environments. But this week the department said seven companies, not one, will be integrated into Impact Levels 6 and 7 environments — the cloud security tiers tied to classified data and missions. The Pentagon fram(cybernews.com)nd stronger decision-making for warfighters. (openai.com) ### Why do Impact Levels 6 and 7 matter? Because this is where “enterprise software” turns into military infrastructure. Impact Level 6 is tied to secret information, and the Pentagon’s cloud contracting stack is built to reach classified levels up through top secret. So these deals are not about a chatbot helping draft emails. They are about letting commercial models and AI toolin(openai.com)s. (dodcio.defense.gov) ### Why is Anthropic missing? Not because Anthropic was absent from defense work before. It had a two-year prototype agreement with the DoD, announced July 14, 2025, with a $200 million ceiling. Anthropic also said Claude had already been deployed on classified networks and inside defense workflows. The break came after negotiations over use restrictions. (dodcio.defense.gov) accept that position. (anthropic.com) ### What happened in March? The dispute escalated fast. Anthropic said on February 27 that Defense Secretary Pete Hegseth was directing the department to designate the company a supply-chain risk. On March 4, Anthropic said it received a formal letter confirming that designation. Anthropic has argued the move is legally unsound(anthropic.com)o punish a U.S. company over contract terms. (anthropic.com) ### So is this really about safety terms? Basically, yes — but also leverage. OpenAI’s February 28 post said its agreement explicitly bars domestic surveillance of U.S. persons and bars use by Pentagon intelligence agencies like the NSA unless a new agreement is reached. Anthropic said its own red lines were narrower in one sense and harder in another: no mass domestic surveillance(anthropic.com)” than “who gets to set the ceiling on lawful military use.” (openai.com) ### Why add seven vendors instead of one? Because the Pentagon does not want a single-lab dependency in a market moving this fast. Its January 2026 AI strategy talks openly about military AI dominance, and broader defense modernization documents describe an “AI-first warfighting force.” Multi-vendor access also lets the department mix models, cloud providers, chips, and specialized(openai.com)y’s stack. That is procurement strategy as much as technology strategy. (media.defense.gov) ### What’s the bottom line? This is the Pentagon telling the AI industry that classified access goes to companies willing to meet defense terms at scale. Anthropic’s exclusion matters because it shows the government is not just buying models — it is setting the rules of entry for the most valuable national-security AI market in Washington. (militarytimes.com)