Dario Amodei 稱 Anthropic 為安全考量延後四個月開放 Mythos 級模型

Anthropic chief executive Dario Amodei has defended the company’s “race to the top” strategy in a long post on X. He was replying to Nathan Calvin, chief strategy officer at the advocacy group Encode AI. Amodei listed nine decisions he said cost Anthropic commercially but made AI safer. He also said the company could have handled some of Calvin’s points better.

Calvin’s criticism followed Anthropic’s response to the backlash over its leaked AI safety bill. On 19 October, he wrote that the company’s answer was “still not really engaging with the heart” of the issue. Amodei posted his reply early on Wednesday. It has been viewed more than 800,000 times.

“Our strategy has always been what I’ve called the ‘race to the top’,” Amodei wrote. “By showing how it’s possible to build the technology responsibly, we can raise the bar for the entire industry.”

He said that strategy only works if the company practising it is at the frontier. He said he was “proud of our track record” and stood by it.

Thank you for this, @_NathanCalvin. I appreciate and am a fan of the investigative work you’ve done on behalf of AI safety and believe your comments are coming from a constructive, honest place.

Dario Amodei (@DarioAmodei) October 21, 2026

The nine decisions

Amodei said Anthropic first had its Mythos-class models in February. It did not make them generally available until June, when stronger cyber safeguards were in place. Until then, only a small group of partners working on cyber defence could use them. Earlier this month, Anthropic said China’s GLM-5.3 nearly matches Mythos at cyber exploits.

“We likely sacrificed a significant chunk of a several-month model lead,” Amodei wrote.

He said Anthropic’s biological safeguards are strict enough that biologists joke about them, and that rivals sometimes sell against the company on this point. He said Anthropic had seen and blocked possible attempts to use its models for bioweapons.

The list also included Anthropic’s role in California’s SB 53 safety law and his call in June for the government to have the power to block model releases. It included the company’s support for export controls on chips to China. Amodei said Anthropic was, as far as he knew, the only frontier company to restrict sales to companies controlled by the People’s Republic of China. He said distillation, training one model on another model’s answers, is no major commercial threat but raises security concerns.

The last item was the new public benefit fund, which will hold 5% of Anthropic’s shares, and possibly more over time.

What Amodei conceded

Amodei said Anthropic could have handled three of Calvin’s points better. He put them down to “rhetoric or policy communication that wasn’t carefully written”, in a company of more than 5,000 people.

“I am of course the CEO, and I take responsibility,” Amodei wrote.

He called Calvin’s point about changes to Anthropic’s Responsible Scaling Policy, its internal safety rulebook, “the most substantive item”. He said the old version was too rigid. On running a separate compliance framework next to that policy, he said he would “look into it further”.

Amodei also said Anthropic is nearing the point where AI moves too fast for safety work to keep up. That is shaping its thinking on “pacing”, he wrote, and he would “have more to say soon”.

The pushback

Sigal Samuel, who covers AI for Vox, asked Amodei on X for evidence that the strategy works.

“From the outside, it looks like Anthropic has done more on net to drive a race to the bottom,” Samuel wrote.