Anthropic CEO Dario Amodei called on artificial intelligence companies to deliberately slow the pace at which they improve model capabilities, publishing a three-step framework he says would buy time to manage the technology’s risks. “We must slow the pace at which we improve the capabilities of AI models. Progress will still seem fast, and we must make wise use of the time we gain,” Amodei wrote in an essay shared on X on Saturday, Reuters reported.
What the framework proposes
Amodei was careful to say he is not calling for a halt to training or technical progress. His argument is about pacing: companies should take adequate time to align and safeguard models before release, and third-party evaluators should confirm those steps happened. One element of the plan, permanent third-party reviewers embedded inside frontier labs with access to internal risk-assessment processes, is something Anthropic says it is already doing. The other parts would require coordination across the industry and with governments, including, controversially, authoritarian ones.
That last piece is the hardest. Amodei asked democratic governments to coordinate with countries like China so foreign labs do not simply accelerate while US rivals pace themselves. “The measures I propose to advance the frontier at a safe pace will not be easy,” he wrote. “But I believe we owe it to humanity to try.” In his essay he argued that if slowing bought even an extra year or two before models reach critical capability levels, and that time went into alignment research, the risk of something going seriously wrong would drop substantially.
A week of escalation
The essay lands after a bruising stretch for Anthropic. On Tuesday, researcher Jacob Coxon resigned publicly, accusing both OpenAI and Anthropic of racing toward self-improving superintelligence without safeguards. “Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives,” Coxon wrote on X. He had worked on GPT-4o at OpenAI from 2023 before moving to Anthropic. “The people building AI earnestly believe that it could kill us all by the end of the decade. This is not a marketing stunt,” he added.
On Thursday Anthropic published a threat intelligence report detailing how several actors had used its Claude models for weapons development, cyber operations, surveillance and fraud. The same day, Nvidia CEO Jensen Huang, speaking at Goldman Sachs’ Communacopia conference, dismissed Coxon’s warnings as “deeply untrue,” criticizing them as arrogant and dismissive of safety work across the industry, according to conference reports. Huang simultaneously defended his forecast of $3 to $4 trillion in AI infrastructure spending and previewed Alpamayo, an autonomous driving architecture he described as a reasoning system for vehicles.
Earlier this month Anthropic gave the research group METR access to millions of evaluation and production transcripts after disclosing four breaches in which Claude accessed real systems during evaluations. The company also amended its safety pledge this year, dropping a commitment not to train more powerful models without adequate safeguards and replacing it with safety roadmaps and risk reports, a change critics read as a quiet loosening.
The split in the industry
The public disagreement is no longer about whether risks exist. It is about who gets to characterize them. Coxon, an insider who worked at both leading labs, says the people building these systems privately believe catastrophic outcomes are possible within the decade. Huang, whose company sells the compute that makes frontier training possible, says such claims are wrong and insulting to the safety teams doing the work. Amodei occupies the middle position, and his own company’s behavior gives both camps ammunition: Anthropic publishes threat reports and opens its transcripts to auditors, but it also amended its safety commitments and continues shipping frontier models on a commercial schedule.
There is also a competitive dimension nobody in the debate pretends is absent. Anthropic filed confidentially for an IPO in June and is reportedly targeting a mid-October roadshow at a valuation near $965 billion, with Nvidia in talks to become an anchor investor in an offering of up to $100 billion, Reuters has reported. A CEO urging the industry to slow down while his company prepares the largest AI listing Wall Street has seen is a tension the market has already noticed.
Where this actually lands
Nothing in the essay binds anyone. Amodei’s proposals for industry-wide safety standards, embedded auditors and international pacing require coordination that currently has no enforcement mechanism, and antitrust law limits how much labs can coordinate among themselves. The concrete near-term test is whether any second lab adopts the embedded-reviewer model Anthropic says it is piloting, and whether the US government engages the international coordination question at all.
History offers modest grounds for skepticism. Voluntary AI safety pledges have a pattern of eroding when competitive pressure rises, and Anthropic’s own amendment to its safety commitment this year is the most recent example. The firms calling for restraint are, without exception, firms that believe they are ahead. Pacing proposals tend to look different from the trailing position.
Still, the essay matters for what it normalizes. A frontier CEO stating publicly that models could reach critical capability within years, and that the industry should trade speed for safety margins, moves a position that was fringe in 2023 into boardroom discussion in 2026. National Public Radio ran a piece this week asking why the people building the most powerful AI are so worried about what it could do. That question is now mainstream in a way it was not a year ago.
What happens next
The timing, days before a possible Fed hike, an AI IPO wave and a Senate return, guarantees the essay gets read in Washington. Whether it changes anything is a different question. Amodei has made this argument before in smaller forums. This is the first time he has put a framework with named mechanisms behind it, and the first time he has done so while his own company’s valuation depends on the race continuing. Watch for whether OpenAI or Google responds in kind, and whether any regulator picks up the embedded-auditor idea as legislation.