Amodei says Anthropic's safety delay cost it a model lead

A buyer choosing between model providers now has one CEO publicly saying his safety process cost his company a competitive edge. Dario Amodei posted a long reply to AI safety researcher Nathan Calvin on Tuesday, defending Anthropic's handling of its Mythos-class models and conceding, in his own words, that the delay "likely sacrificed a significant chunk of a several-month model lead." The X thread is the primary source; The New Stack reported it the next day and asked Anthropic for additional context.

The sequence Amodei is defending: Anthropic first released Claude Mythos Preview in April through Project Glasswing, giving selected partners (Amazon, Apple, Microsoft, CrowdStrike, and the Linux Foundation among them) access to the model's vulnerability-discovery capabilities for defensive work, while keeping it out of the hands of ordinary developers. On 2 June the program expanded to roughly 150 additional organisations across more than 15 countries. The gap, from the first restricted release to the broader expansion, is the period Amodei is calling a model-lead cost.

What Calvin argued

Calvin posted his criticism on Monday. He argues that Anthropic increasingly resembles a company racing to build superintelligence while weakening its safety commitments. His case names the company's more than $500 billion in computing deals, its growing use of AI to develop new models, and specific changes to its Responsible Scaling Policy. "When Anthropic strays and makes mistakes, why do the mistakes systematically tend to guard and favor Anthropic's interests?" Calvin writes in his thread.

The policy change at the centre of the argument is Anthropic's February revision to version 3.0 of its Responsible Scaling Policy, since updated to version 3.4 (effective 8 July). The revision replaces earlier requirements that Amodei now describes as too rigid with an approach that emphasises transparency, risk reporting, and independent evaluation. The February change also removes a 2023 commitment never to train an AI system unless Anthropic could guarantee in advance that its safety measures were adequate. Calvin argues this weakens Anthropic's unilateral commitment to slow or stop development.

What Anthropic is saying

Anthropic told The New Stack the Mythos delay follows a commitment in its Responsible Scaling Policy to require a strong argument that catastrophic risk is contained, and to delay deployment "as far as needed until it no longer believes it holds a significant lead." Amodei frames the overall approach as a "race to the top," arguing that Anthropic can shape industry safety standards more effectively by staying at the technological frontier than by stepping back from it.

Jared Kaplan, Anthropic's chief science officer, gave Time magazine the plain version of the trade-off in a February story. "We felt that it wouldn't actually help anyone for us to stop training AI models," Kaplan said, adding that unilateral commitments did not make sense "if competitors are blazing ahead." Amodei concedes Anthropic could have handled three of Calvin's points better, including its handling of a draft bill, and calls the Responsible Scaling Policy critique "the most substantive item" in Calvin's post.

A buyer reading this picks up one new fact: Anthropic's own CEO has stated on the record that a safety process cost the company a model lead measured in months. A buyer who wants a provider that will not pause for safety reasons now has that signal. A buyer who wants the opposite also has it. Both get to decide against public evidence rather than against a brochure.

Source

Dario Amodei on X: reply to Nathan Calvin, 20 October 2026. Press coverage: The New Stack: Amodei says Anthropic's safety delay may have cost it an AI lead.