The seven contracts the Pentagon signed last week ran the table on the American AI industry except for one company. OpenAI got one. Google got one. Microsoft got one. Amazon got one. Nvidia got one. SpaceX got one. The seventh went to Reflection, the open-weights upstart that did not exist eighteen months ago. Anthropic — the lab whose Constitutional AI papers the Defense Innovation Unit cited as recently as 2024 — got nothing. [1]

The omission is not bureaucratic drift. A senior Pentagon official told Government Executive on Wednesday that the department "will never again rely on a single AI provider," language that reads at first like vendor diversification and at second like the formal version of the doctrine the building already operates: Anthropic is not one of the providers being diversified among. [2] The paper's Saturday account framed this as a two-track government — Pentagon excludes, civilian agencies route around it. The seven-contract list confirms the architecture.

What makes the week strange is what the West Wing was doing while the Pentagon was finishing the contracts. Axios reported April 29 that the White House is drafting an executive order that would unwind the OMB directive instructing civilian agencies not to procure Anthropic models. [3] Nextgov, which has the most granular reporting on the drafting process, identified at least two table-reads of the EO inside the policy-coordinating bodies that walk drafts toward the Resolute Desk. [4]

So the executive branch is, in writing, operating two contradictory AI-procurement doctrines at the same time. The Pentagon is hardening the Anthropic exclusion. The civilian White House is preparing to lift it. Trump told CNBC's Becky Quick on Tuesday that Anthropic is "shaping up." [5] He did not say what shape.

The Pentagon track is the more visible one because contracts are public and because the seven names fit on a chyron. The granular money — most of these are $200 million ceilings on multi-year IDIQ awards, not seven nine-figure checks — matters less than the fact that the building has now built its short-list without the company that was, in 2024, considered the only frontier lab that took national-security alignment seriously. The Anthropic Researcher Access Program inside the National Security Commission on AI was, at one point, the model for how a frontier lab and the federal government were supposed to talk. The model is now a counterfactual.

The reasons the Pentagon walked are not, by any DoD official's on-record telling, technical. Claude's benchmark performance against the OpenAI and Google frontier models is comparable. Anthropic's Computer Use API, demoed for the Defense Innovation Unit last fall, was, by people who saw it, the most capable agentic interface among the labs. What the Pentagon could not square was Anthropic's published policy of refusing certain weapons-adjacent use cases without case-by-case sign-off. A frontier model that requires Pentagon program managers to ask permission of a San Francisco trust-and-safety team is, by the building's logic, not a Pentagon model.

The civilian-side EO drafting is the inverse story. The reasoning inside OMB and OSTP, per Nextgov, is that locking civilian agencies out of Anthropic creates exactly the problem the Trump AI policy was supposed to prevent: federal agencies stuck on a narrower vendor menu than the private market. The HHS Office of the Chief AI Officer, Treasury's data-science group, the Department of Energy national labs — these are the customers that an EO walk-back would be addressed to.

What's missing in the Pentagon coverage is the pricing. The seven contracts together come to under $2 billion in obligated ceiling — small by Pentagon procurement standards, large by AI-vendor standards. The Reflection award is the surprising one: a fourteen-month-old open-weights startup beat Anthropic for the same slot. People close to the procurement say Reflection's open-weights posture mattered more than its capability scores. The Pentagon wanted to be able to fine-tune the models inside its own SCIFs without a vendor in the loop. That is the doctrine the seven-vendor list expresses, even if no memo says so out loud.

What's missing in the EO coverage is the timing. The drafts have been in table-read for at least two weeks. They have not surfaced as text. People who have seen them describe a one-page directive, light on procurement specifics, heavy on the language of "competitive neutrality." The version Nextgov was briefed on includes a recital paragraph distinguishing civilian from defense procurement, which would let the Pentagon's exclusion stand while the civilian agencies regain access. [4]

That is the two-track government in writing. The Pentagon gets to keep its short-list. The civilian agencies get to add Anthropic back. The framework lets both things be true at once. Whether it survives contact with the Hill — Senator Cotton's office is already on record that any EO walking back the OMB directive will draw a hearing — is a different question. But the structural fact, as of Sunday, is that the same executive branch that is drafting Anthropic's reentry is the executive branch that just excluded the company from the largest AI-procurement round in American history.

Anthropic's San Francisco public posture has been silence. The company's Mythos Preview release, scheduled for late May, is the next Anthropic surface that will tell the federal customer something. Whether the Mythos guardrails are tightened or loosened from the Claude posture the Pentagon found unworkable is the variable nobody outside Pier 33 knows.

-- THEO KAPLAN, San Francisco