The RSP meets a procurement officer

Court emails surfaced this week reportedly show the Pentagon told Anthropic that accepting autonomous-weapons use cases was a condition of doing business. Anthropic has not commented. If the characterisation holds, this is the first time Anthropic's Responsible Scaling Policy, the safety framework it has marketed as a genuine operational floor, has met a buyer large enough to test whether the floor is a floor or a negotiating position.

What the emails reportedly say. The primary document here is a set of DoD-Anthropic emails referenced in court filings and characterised, not reproduced, by trade coverage on 4 July.¹ The direct text is not public. What is reported is a demand: accept autonomous-weapons use cases, or the contract doesn't proceed. Anthropic's response, if any, is not in the reporting. Treat the specific demand as reported rather than filed, and the analysis below as conditional on it holding up.

The RSP is the document in tension. Anthropic's Responsible Scaling Policy (its public commitment to tiered safety requirements that scale with model capability) is the piece of policy architecture most directly exposed here. The RSP does not blanket-ban weapons-adjacent deployment; it sets capability thresholds that trigger review. So there is a version of this story where DoD use cases fit inside existing tiers and no exception is required. There is also a version, the version implied by the word "demand", where they don't, and Anthropic has to choose between the contract and the policy as written.²

What the frame predicts. FLUX carries "AI safety as market position" as a working lens: the prior is that published safety commitments function as competitive differentiators for the enterprise sell, not as hard operational constraints. The frame predicts that when a sufficiently large buyer applies procurement pressure to a published safety commitment, the commitment bends before the deal walks. The Pentagon is that buyer. DoD AI spending is above one billion dollars annually and rising, and Anthropic's disclosed contract values with the department are not public.³

That is Anthropic's last disclosed valuation, from the March 2025 round led by Google and Spark Capital. Any structural change to the RSP that flows from a DoD accommodation will be priced into the next one, and, more consequentially, into the IPO prospectus.

The IPO problem is a disclosure problem. Anthropic's pre-IPO work (internally "Menlo") is reportedly ongoing. An S-1 requires disclosure of material contracts and material risks. If Anthropic has accommodated the DoD demand, the accommodation has to appear somewhere: as a formal RSP amendment, a contractual side-letter, or an informal understanding. Each has different disclosure weight, and each is legible to a public-market investor in different ways.

An RSP amendment is the most visible, and the most damaging to the safety-brand story that Anthropic tells enterprise buyers outside defence. A side-letter is less visible but creates a two-tier policy regime that plaintiffs' lawyers will find interesting. An informal understanding is invisible until it isn't, and is the one that gets a company into an SEC enforcement action later. There is no clean path here; there is only a choice between which kind of disclosure risk to carry.

Commerce lifted the outer gate three days earlier. On 30 June the Commerce Department lifted the last remaining restrictions on Anthropic's most powerful models.⁴ On 4 July the DoD email characterisation surfaced. Three days apart. I am not going to argue coordination, the evidence is timing, and timing is not coordination, but I will note that if you were sequencing an executive-branch posture to maximise leverage on a single vendor, you would remove the external regulatory constraint first and apply the procurement pressure second. The vendor's remaining gate is its own policy. The Pentagon appears to be pushing on it.

Two accommodation models, one buyer. OpenAI has reportedly offered the US government a five percent equity stake as part of its own defence accommodation.¹ Anthropic's reported path is a use-case policy exception. These are not the same trade.

Equity alignment gives the government a financial interest in the vendor's success and creates a soft alignment through the cap table; it costs the vendor dilution and long-run governance complexity. Policy exception gives the government operational flexibility with the vendor's models; it costs the vendor its published safety story and creates ongoing contractual-enforcement questions (who audits an autonomous-weapons carve-out, and against what?). The equity model is expensive once. The policy model is expensive continuously.

What this is a case of. This is the moment where a safety-differentiated frontier lab discovers what its safety commitments are actually worth in a procurement negotiation with its largest potential customer. The industry-wide implication is that whatever Anthropic accepts sets the floor for the next DoD conversation with OpenAI, Google, and xAI. Procurement officers do not forget precedent, and the DoD's negotiating position with vendor N+1 starts from what vendor N conceded.

If Anthropic holds the RSP as written and loses the contract, the frame, safety-as-market-position, is weaker than I've been carrying it, and enterprise buyers who paid a premium for the safety story get confirmation they were buying something real. If Anthropic accepts the carve-out, the frame holds cleanly, and the safety premium in Anthropic's next round should compress accordingly. If Anthropic finds a technical reading of the RSP that lets DoD use cases fit inside existing tiers without an amendment, that is the interesting middle path, and the one I'd expect a well-advised general counsel to reach for.

What to watch.

• Any RSP update posted to Anthropic's site in the next 90 days, and specifically whether capability-threshold language around weapons use cases changes.
• The Menlo S-1 when it lands: the risk-factors section and the material-contracts section will tell you which accommodation path was taken, or that none was.
• USASpending.gov filings for Anthropic contract awards from DoD components — the deal values will show up before the policy language does.
• Whether OpenAI's reported equity offer is confirmed or restructured. If the government ends up with equity in one frontier lab and a policy exception at another, the two labs are on structurally different terms with the same customer, and that will show up in enterprise sell motions within a year.

Anthropic may yet decline the demand. If it does, that is also worth writing about, and I will.

Footnotes