Anthropic waited until June to give general access to Mythos-class models it had first in February, Dario Amodei said. He presented the delay as a safety decision with a commercial cost. His explanation came after Nathan Calvin criticised the company’s response to its leaked AI safety bill. The exchange gives readers a view of both Anthropic’s stated safeguards and the questions still open.
What prompted Amodei to respond?
Nathan Calvin, chief strategy officer at Encode AI, criticised Anthropic’s answer to backlash over the leaked safety bill. On 19 October, he wrote that the response did not address the heart of the issue. Amodei replied in a long post on X early on Wednesday. According to the source report, the post had been viewed more than 800,000 times.
Amodei thanked Calvin for his work and described the criticism as constructive. He also said Anthropic could have handled three of Calvin’s points better. He linked those shortcomings to rhetoric or policy communication that had not been carefully written. He said the company had more than 5,000 people and accepted responsibility as CEO.
What does “race to the top” mean here?
Amodei used the phrase for a strategy of building AI responsibly while remaining at the frontier. He argued that demonstrating responsible development could raise expectations across the industry. The claim is his account of Anthropic’s approach, rather than an independent finding about its effects. He said he was proud of the company’s record and stood by it.
The idea matters to founders and investors because safety choices can affect access and competition. Amodei’s example was that Anthropic had Mythos-class models in February but waited until June for general availability. Before then, only a small group of partners working on cyber defence could use them. He said stronger cyber safeguards were in place by the time broader access began.
What costs did Anthropic say it accepted?
Amodei said the delay likely cost Anthropic a significant part of a lead lasting several months. The source does not quantify that cost or define the lead in financial terms. Earlier this month, Anthropic said China’s GLM-5.3 nearly matched Mythos on cyber exploits. That statement provides company context but does not independently establish the relative capabilities of the models.
His list also covered biological safeguards, which he described as strict. He said biologists joked about them, rivals sometimes sold against Anthropic on that basis, and the company had blocked possible efforts to use its models for bioweapons. The source excerpt offers no detail about those efforts. Amodei also cited support for export controls on chips to China and restrictions on sales to companies controlled by the People’s Republic of China.
Which policy positions did he include?
Amodei referred to Anthropic’s role in California SB 53 and his June call for government power to block model releases. He also said distillation was not a major commercial threat but raised security concerns. Distillation means training one model using another model’s answers. These examples broaden his argument beyond product safeguards to public policy and how AI systems are developed.
The final item was a new public benefit fund set to hold 5% of Anthropic’s shares, possibly more over time. The source provides no valuation or other financial terms. Amodei identified changes to Anthropic Responsible Scaling Policy as Calvin’s most substantive criticism. The excerpt does not describe those changes, leaving readers without enough detail to judge the disagreement fully.
Anthropic’s claims describe safeguards and commitments that the company says can carry commercial costs. Amodei has also acknowledged weaknesses in how some policy points were communicated. The available account leaves the details of the disputed policy changes and two other criticisms unspecified. Further explanation from Anthropic would clarify what changed and how the company plans to address Calvin’s concerns.