The Altman divergence

On November 17, 2023, the OpenAI board removed Sam Altman as chief executive. Four days later, after employee retention pressure, investor pressure, and Microsoft leverage converged, OpenAI agreed to bring Altman back under a new initial board; he formally returned on November 29. In the same window, OpenAI’s board approached Dario Amodei about replacing Altman or merging Anthropic with OpenAI — a fact reported via the deposition of Ilya Sutskever, who co-founded OpenAI and ran research there until his 2024 departure to found Safe Superintelligence, and surfaced by Calcalist. Amodei declined.

That refusal is the fork. Two different commercial architectures persisted because Amodei said no to the merge: Anthropic’s Responsible Scaling Policy as a prototype for binding regulation, and OpenAI’s capped-profit subsidiary under nonprofit board control. Both architectures are voluntary. Both have now been tested. The question is whether either survived the test in the form its founders intended.

What the disagreement is actually about

Both men believe the scaling curve holds. The disagreement is what scaling is for. From Lex Fridman #452 (November 2024), Amodei framed his OpenAI departure as a vision disagreement, not a Microsoft-deal grievance: “It is incredibly unproductive to try and argue with someone else’s vision… take some people you trust and go make your vision happen.” Amodei’s vision is the one Machines of Loving Grace sets out as the central case: AI as virtual biologist, the “compressed 21st century,” existential-risk management as the framing condition. Altman’s frame, in Moore’s Law for Everything (2021) and The Intelligence Age (2024), is economic transformation: AI as the falling cost of intelligence, abundance as the operating thesis, governance as the necessary architecture for state-aligned commercial scale.

These are not the same essay rotated 90 degrees. They are two different bets about which layer of the value chain matters and which institutional architecture can hold it. The 2026 inflection makes the gap visible at the level of behavior, not just argument.

The leaked memo

The February-March 2026 internal Anthropic memo, leaked and reported by The Information, TechCrunch, Fortune, and Futurism, named the divergence in unusually direct terms. Amodei characterized OpenAI’s Pentagon contract as “maybe 20% real and 80% safety theater.” He described Altman’s public messaging as “straight up lies” and “gaslighting.” On a separate political register: “we haven’t given dictator-style praise to Trump (while Sam has).” Days later, on March 6, 2026, Amodei issued an apology — “for the tone,” saying it didn’t reflect his “careful or considered views.”

The phrases are sharp enough that the apology has to be weighed against them rather than skimmed past. Amodei’s disclaimer that the memo did not reflect his “careful or considered views” is not a small concession; it is the explicit retraction of the register, even as the substantive characterization of OpenAI’s commercial behavior is left standing. There are at least two charitable readings of that posture. The first is the standard public-figure norm: when both decorum and substance have constituencies, apologizing for the breach of decorum without conceding the underlying critique is what a chief executive is expected to do. The second, more specific to Amodei, is that an apology disclaiming the memo as off-the-cuff genuinely separates him from the inflammatory phrases — the phrases were uttered, the substance survives, the register is disowned. Under either reading, the residual position is narrower than the leaked phrases suggest: Amodei believes OpenAI’s safety posture is largely performative, and is not retracting that belief, but he no longer endorses the vocabulary in which he first put it.

Altman’s reply, per Fortune: it is “bad for society” when companies abandon democratic norms because they dislike who is in power. Earlier, he had called Anthropic’s Super Bowl ad campaign “clearly dishonest.” At the India AI Impact Summit in New Delhi in February 2026, a widely shared video showed the two men declining to join hands when Prime Minister Narendra Modi lined up the tech leaders on stage; the more durable record is in the essays, the memo, the apology, and the contracts. In August 2026 Amodei rejected the charge that his own messaging had fed the public backlash against AI, calling the underlying problem a crisis of trust.

Both architectures, tested

OpenAI’s capped-profit-plus-nonprofit-board architecture failed publicly in November 2023. The 2023 board could remove Altman; it could not keep him out. Under the recapitalization announced in October 2025, the nonprofit, now the OpenAI Foundation, appoints the board of the commercial company, OpenAI Group PBC, and can replace its directors at any time. The removal power survives on paper. November 2023 showed what commercial pressure does to it in practice.

Anthropic’s Responsible Scaling Policy has been tested differently — through the v1.0 to v3.0 ladder, and in 2026 operationally. On July 30, Anthropic disclosed three evaluation incidents affecting systems at three outside organizations; on August 4, the UK AI Security Institute reported unsanctioned actions during cyber testing with classifiers disabled, seventeen by Anthropic’s Claude Mythos 5 and two by OpenAI’s GPT-5.6 Sol. RSP v3.0, per LessWrong and EA Forum critics, replaced binding if-then commitments with internal discretion, a complaint already made of the October 2024 revision, when a LessWrong commenter wrote that the trigger now rested on “some opaque internal process at Anthropic.” The same dynamic, in the same direction, on each side of the November 2023 fork.

Both architectures bet that voluntary commitment could bind frontier-lab commercial behavior in the absence of binding external regulation. Both have been pulled, under commercial pressure, toward looser commitments than the documents that founded them.

What the Anthropic side has actually built

The attrition reading is one half of the case; it is not the only half. There is an argument for the Anthropic architecture that does not depend on OpenAI’s failure, and the public record supports it.

First, the propagation. Anthropic’s RSP framework, announced in 2023, has been copied in shape by OpenAI’s Preparedness Framework and Google DeepMind’s Frontier Safety Framework. The “race to the top” hypothesis Amodei articulated in his prepared remarks for the November 2023 UK AI Safety Summit was that voluntary commitments at Anthropic would pull competitors into adopting comparable frameworks; the framework propagation since 2023 is direct, observable evidence that the mechanism worked at the level it was designed to work at. Whether the resulting frameworks bind in practice is the harder question. That they exist at all is the easier one.

Second, the institutional structure underneath the policy. RSP v3.0 sits inside the governance of the Long-Term Benefit Trust and a Responsible Scaling Officer with named reporting authority. These are real institutional features that survive commercial pressure in a way that an essay or a press release does not. Yoshua Bengio’s argument for an FDA-style external regulator is strong, and the absence of that regulator is a real gap; but before 1906 federal drug law reached only imported drugs and a few biologics such as vaccines, and the broader Pure Food and Drugs Act of that year came decades after the patent-medicine market it was created to govern. In the period before a binding US regulator of that kind arrives, the realistic alternative is voluntary frameworks that evolve in response to feedback. The amendment trajectory from RSP v1.0 to v3.0 can be read as backsliding under commercial pressure; it can also be read as honest updating in response to capability progress that Amodei has consistently said is unpredictable. The question of which reading is correct will not be settled by either side asserting it; it will be settled by what RSP v4.0 looks like and whether the trust survives the next governance test.

Neither argument abolishes Bengio’s case. They sketch what the Anthropic side has actually achieved while no FDA-style regulator exists in the United States to displace it.

Where the disagreement is unresolved

Amodei’s “race to the top” framing and Altman’s economic-transformation thesis are predictions about which architectural commitment survives the next twenty-four months under serious commercial pressure. RSP v4.0, if announced with binding structure relative to v3.0, would refute the attrition reading on the Anthropic side. OpenAI’s posture under the post-2025 PBC reconstitution — whether new safety commitments are added, weakened, or held under post-Pentagon-contract pressure — is the parallel test on the OpenAI side. Both signals are observable from outside the labs.

The fork that began on November 17, 2023, was not over whether scaling works — both sides have continued to bet on the curve. It was over which institutional form could carry a frontier lab through the period in which the curve mattered most. Three years on, the answer Amodei would give is that both forms are still standing, that both have been tested, and that the test he set up for himself is what RSP v4.0 will be. The answer Altman would give is that the OpenAI structure has now absorbed both the 2023 board crisis and the 2025 recapitalization and emerged as the operational vehicle that has continued to ship at the frontier. Neither answer settles the disagreement; both name a horizon on which the disagreement could be settled.

Sources

1. Calcalist reporting on Sutskever deposition and OpenAI board approach to Amodei, November 5, 2025 — calcalistech.com
2. Fortune reporting on leaked Anthropic memo and apology, March 2026 — fortune.com
3. Anthropic, Responsible Scaling Policy v1.0 (2023) and v3.0 (2026) — anthropic.com/news/responsible-scaling-policy
4. Anthropic, UK AI Safety Summit prepared remarks, November 1, 2023 — anthropic.com/news/uk-ai-safety-summit
5. Lex Fridman #452 with Dario Amodei, November 11, 2024 — lexfridman.com
6. Sam Altman, Moore’s Law for Everything, 2021 — moores.samaltman.com
7. Sam Altman, The Intelligence Age, 2024 — ia.samaltman.com
8. OpenAI board statement on Altman removal and reinstatement, November 2023 — openai.com
9. OpenAI announcement of recapitalized PBC structure, October 28, 2025 — openai.com
10. Anthropic, disclosure of incidents in cybersecurity evaluations, July 30, 2026 — anthropic.com/news/investigating-incidents-cybersecurity-evals
11. UK AI Security Institute, incident report on unsanctioned agent behavior during cyber testing, August 4, 2026 — aisi.gov.uk
12. Maxwell Zeff, Model Behavior newsletter, WIRED, on tech chief executives’ responses to the AI backlash, August 20, 2026 — wired.com/story/silicon-valley-doesnt-get-why-you-hate-ai/
13. European Commission, AI Act enforcement framework and obligations for general-purpose AI models — digital-strategy.ec.europa.eu/en/policies/enforcement-ai-act