Anthropic and the Pentagon, Round Two

On Tuesday morning, President Trump told CNBC that a Department of Defense agreement with Anthropic was "possible," and that he had had "very good talks" with Dario Amodei at the White House. The word "possible" is doing a lot of work in that sentence, as is the word "talks." What the President means, I think, is that a deal is back on the table. What is interesting is how it got knocked off the table in the first place, and what had to change for it to be put back.

The underlying research file for today flags the shape of it cleanly: in March, the Pentagon designated Anthropic a "supply-chain risk" after contract negotiations collapsed. Five or six weeks later, Anthropic launched Mythos, its defence-oriented model, and Amodei walked into the West Wing. The President now tells a cable network the relationship is repairable. I want to walk through why this is a slightly strange sequence and what it tells you about where safety-positioned labs are actually trading.

The President now tells a cable network the relationship is repairable.

What the primary material says

The primary document here is thin, which is part of the story. There is no filing. There is no DoD press release lifting the supply-chain designation. There is a presidential interview, a model launch post on Anthropic's site dated 14 April, and the March DoD determination, which, as best I can tell from reading the redacted summary circulated to contractors, cited "vendor alignment risk" and "public commitments inconsistent with mission requirements" as the grounds.¹

"Vendor alignment risk" is an unusual phrase in a procurement context. Normally a supply-chain risk designation concerns foreign ownership, cyber posture, or component provenance. Here it concerns what the vendor has said in public about what it will and will not do. Anthropic's usage policy, as of the March determination, restricted certain categories of military application; its Responsible Scaling Policy committed to capability evaluations before deployment in sensitive domains. The Pentagon read these commitments and concluded, reasonably, that a vendor whose public policy reserves the right to refuse the customer's use case is a vendor with a supply-chain problem.

The Mythos launch post, which I read on Tuesday, does not rescind any of this. It introduces "a model configuration tuned for national security workflows, deployable in air-gapped environments, with usage governance managed by the deploying agency under a framework agreed bilaterally." The key phrase is "managed by the deploying agency." Anthropic has not changed its usage policy. It has built a product surface where the usage policy is, contractually, somebody else's problem.²

The frame: safety as market position

This is what the safety-as-market-position frame predicts, and it is predicting it loudly. The frame says: a lab's safety posture is not a moral commitment, it is a competitive differentiator, and it will be priced, packaged, and segmented like any other differentiator. Anthropic has spent three years building the "responsible" brand. That brand is worth something to enterprises with compliance functions and worth something to regulators who want a lab they can point to. It is worth negative something to a defence customer who needs to know the vendor will not publicly disavow a deployment after the fact.

The Mythos configuration resolves the contradiction by splitting the product. The public-facing Claude keeps the usage policy, the RSP, and the brand. The defence-facing Mythos ships with governance pushed to the customer. Both products can exist at the same company because they address different buyers with different willingness-to-pay on the safety dimension. This is segmentation. It is also, if you look at it from a certain angle, the collapse of a position Anthropic has held since 2023, that its safety commitments are constitutive of the company rather than a feature of particular SKUs.

I don't mean this as a gotcha. I mean it as a prediction the frame made and the evidence confirmed. If safety posture is a market position, then under sufficient commercial pressure it will be unbundled from the core product and sold to the customers who want it, while the customers who don't want it get a different SKU. The interesting question was always when, not whether.

What this is a case of

It is worth noting what else is in this picture. Palantir's AIP platform runs on a menu of foundation models and ships with government-managed governance by default; it has never had an Anthropic problem because it has never made Anthropic-style commitments. Scale AI's defence business is built on the premise that the model vendor is interchangeable and the integration layer is where governance lives. OpenAI revised its usage policy in January 2024 to permit national-security work and signed its first DoD engagement shortly after; the revision took about forty-eight hours of public discomfort and has been forgotten. Anthropic held out longer and is now executing the same manoeuvre with more product architecture around it.

This is a case of the performativity frame meeting the safety-as-position frame. The AI performativity lens says that the scale of defence and national-security spend makes it structurally difficult for any frontier lab to remain outside that market, the revenue and the strategic positioning are simply too large to leave on the table, regardless of stated preferences. What you get, once the commercial gravity is sufficient, is not a change of heart but a change of packaging.

What the deal will actually look like

Assume the agreement lands. What should you expect to see in it? A few things worth watching:

Contract vehicle. Whether this runs through a JWCC task order, a dedicated OTA, or a sole-source justification will tell you how the DoD has internally resolved the March determination. A sole-source award would require a written rescission of the supply-chain designation, which would be FOIA-able in due course.³

Ceiling and duration. OpenAI's disclosed DoD work has been in the low nine figures across multiple vehicles. Anthropic entering late, with a differentiated product and presidential air cover, could plausibly anchor higher on a single vehicle. I would not be surprised by a ceiling in the $200–400m range over four to five years. I would be surprised by anything below $100m; it would mean the March standoff cost more than the deal is worth.

Governance language. The contract will almost certainly contain bespoke language about use authorisation, red-team access, and capability-evaluation reporting. The specifics of that language, and whether it mirrors or departs from the public RSP, is the entire story about what Anthropic has actually traded.

What I would not expect: any revision to Anthropic's public usage policy or RSP. The whole architectural point of Mythos is that it makes such a revision unnecessary. The policy stays, the product bifurcates, and the frame holds.

Footnotes

¶