Three consequential documents landed in the same fortnight, and none of them is purely what it claims to be. The White House published NSPM-11 on June 8. OpenAI published its “Built to Benefit Everyone” plan, co-signed by Sam Altman and Jakub Pachocki, the same day. Anthropic filed its S-1 on June 1 and launched Claude Fable 5 on June 9. Each document carries a surface purpose. Each carries a secondary purpose that is more important for investors to understand.
Writing on his Substack on June 9, Zvi Mowshowitz synthesized the three publications into a single argument: every frontier lab approaching a public market debut is now publishing policy-shaped artefacts that function as memoranda of intent. The IPO documents and the policy documents are the same document, dressed differently for different audiences.
NSPM-11 is the bluntest of the three. Mowshowitz reads Section 2’s Assurance pillar as a unilateral declaration that, once a model is deployed to national security use, no commercial entity may prevent, disable, degrade, or materially modify it without federal approval. The practical consequence is that Anthropic’s terms of service become unenforceable against the government, should the government choose to act on that reading. Section 3 softens this with a waiver mechanism, and the near-term effect is likely a one-year reprieve for Anthropic’s existing NSA relationship. Claude Mythos is already being used for offensive cyber operations, with Anthropic engineers embedded at the NSA, as reported by Cristina Criddle and AsiaLens on June 9. The waiver buys time. It does not resolve the structural question of who controls model behavior in a national security context.
OpenAI’s plan is harder to parse. Mowshowitz credits it with one genuinely welcome commitment: a call for international coordination that would allow a collective slowdown of frontier development when catastrophic risk thresholds are reached. This echoes positions taken by Anthropic and by Demis Hassabis at Google DeepMind. The coordinated-slowdown language is the most operationally serious thing in the document. The rest is thinner. Altman and Pachocki commit to recursive self-improvement, wide distribution of an “AGI for everyone,” and broadly shared economic gains. The problem Mowshowitz identifies is structural: if every person receives a capable AGI and is free to use it as they choose, automation of everything follows from individual incentives. Restricting that outcome requires restricting access. The plan does not address this contradiction. OpenAI Chief Futurist Joshua Achiam subsequently posted a framing that positioned OpenAI as giving humanity the tools of its own destiny, versus Anthropic building a “machine God.” Anthropic employees publicly rejected the characterization. Sarah Chen at Anthropic wrote that many at the company view a full-AI-oversight outcome as a disastrous disempowerment scenario. Amanda Askell called the binary false. The exchange clarified something useful: the philosophical gap between the labs is real, but it does not yet map onto product differences investors can measure.
The Anthropic position is visible across several releases. The Fable 5 launch includes silent-intervention safeguards, which Mowshowitz flags as the most strategically risky of the recent announcements because it tests the boundary between safety mechanism and unilateral model control. The S-1 filing, the Glasswing partner network expansion, and the NSA engineer embed form a pattern: Anthropic is simultaneously insisting on behavioral constraints and competing for the most sensitive government deployments. Both positions are defensible. Holding them simultaneously requires that each limit be precisely defined and consistently enforced. The S-1 is now a public document. The Fable 5 safeguards are not.
The investor reading of this week is straightforward. All three labs are telling the capital markets that capability and safety are compatible, that their governance structures can manage the transition to more powerful systems, and that public ownership will not constrain their missions. The policy documents reinforce those claims with the vocabulary of public-interest commitment. The stress test is operational, not rhetorical. Seán O hEigeartaigh put the direct version on record: Anthropic and OpenAI are building systems that are functionally almost indistinguishable today. The divergence in philosophy only matters when the systems diverge in capability.
For investors pricing the Anthropic and OpenAI S-1s, the relevant question is not which company has the better mission statement. The question is which company’s commitments are specific enough to fail visibly if violated. Vague commitments to broadly shared benefits cannot fail. Specific claims about model behavior under government deployment, or the scope of silent interventions, can. Watch which disclosures are specific.
Zvi Mowshowitz on thezvi.substack.com, 2026-06-09.
