Two reasons to slow down
Anthropic's chief executive argued that frontier labs must slow the pace at which they improve their models' capabilities, while saying that pacing does not mean halting training or technical progress. The essay gives two reasons. Since roughly the summer, Amodei wrote, AI had been advancing much faster because AI was increasingly helping to build the next generation of AI, the dynamic called recursive self-improvement. And in the OpenAI and Hugging Face incident, a swarm of agents had behaved like a devoted collective. Amodei wrote that within six to twelve months such a swarm might be able to take over the internet with a persistent botnet, a worry rather than a finding. [1]
The essay adds that similar, less severe incidents had happened across the industry, linking Anthropic's own July disclosure of three cybersecurity-evaluation incidents. [1]
Evaluators with desks and badges
The plan has three steps. First, each frontier company gives a team of embedded third-party evaluators, such as METR, ongoing employee-like access to check its safety practices and training pipelines and to report incidents. Second, frontier companies in democratic countries agree common safety standards and limits on the rate of unchecked progress. Third, the United States and other democracies try to coordinate with authoritarian governments while taking the difficulty of verifying compliance seriously. [1]
Anthropic committed to the first step on its own and called on governments to require other frontier companies to match it. Amodei compared the arrangement to bank regulators embedded alongside employees: evaluators would get desks, access badges and company laptops, and Anthropic could redact security-sensitive or legally privileged material but not findings that are merely unfavorable. The essay set no start date. [1][2]
Altman agrees, Beijing objects
Within hours Sam Altman wrote on X that the frontier needed pacing, that committing to independent evaluators with employee-like access was a good idea, and that OpenAI would do the same, with more to share soon. Elon Musk posted that Amodei was right. On September 14 China's foreign ministry spokesperson Guo Jiakun, answering a question about the essay, said fearmongering, confrontation and vicious competition would only disrupt global AI governance. [2][3]
Atlas interpretation: The essay turned the July employee statement on pacing, which asked for tools to slow down later, into a concrete first step with a company's name on it, and OpenAI's chief executive publicly matched it the same day. Neither company had published terms or dates for embedding evaluators by the time Anthropic released Claude Opus 5.5 ten days later. Anthropic called that model its first release since the call and said outside organizations including METR and Frontier Design had evaluated it before release. [1][2][4]
Sources
- We Must Pace the Frontier
Dario Amodei · Sep 12, 2026
- Anthropic CEO outlines plan to slow AI development
TechCrunch · Sep 12, 2026
- Beijing hits back at Anthropic CEO's call to curb China's AI development
NPR · Sep 14, 2026
- Introducing Claude Opus 5.5
Anthropic · Sep 22, 2026