Anthropic’s Plan to Pace Frontier AI Raises a Hard Question: Can the Race Be Slowed?

Introduction

Should frontier AI companies voluntarily slow the race to build more capable systems? Anthropic CEO Dario Amodei has offered one of the clearest answers so far. In a recent blog post, he called for companies to “pace the frontier”: not a halt to research, but a reduction in the speed of unchecked capability improvements so that safety testing, governance, and incident response can keep up.

OpenAI CEO Sam Altman agreed, saying the issue has been a major topic of discussion at OpenAI in recent weeks. SpaceX CEO Elon Musk also endorsed Amodei’s position.

Three proposed mechanisms

Amodei’s plan has three broad elements:

• Embedded independent evaluators. Anthropic is unilaterally committing to host evaluators from outside organizations such as METR. They would receive company credentials, workspace, equipment, and access broadly comparable to internal risk teams, subject to legal and contractual limits. Their role would include checking whether safety and pacing commitments are being followed and helping ensure that incidents are reported. Altman said OpenAI would do the same and provide more details later.
• Coordination among democratic countries. Leading companies should develop shared safety standards and discuss limits on the rate of unchecked progress. Amodei recognized that coordination between competitors could create antitrust concerns, and suggested that the U.S. government mediate or provide a narrow waiver for certain safety discussions.
• Global coordination on clear red lines. The United States and its allies should try to cooperate with other governments, including China, where possible. Even a broad agreement may be unrealistic, but governments could potentially prohibit narrowly defined, obviously dangerous uses, such as using AI to produce biological weapons or enabling users to do so.

Why the call is emerging now

Amodei pointed to two developments that changed his view: a hacking incident involving OpenAI and Hugging Face, and the perception that AI capabilities have been advancing dramatically faster in recent months. He placed particular emphasis on models’ growing ability to build the next generation of AI systems, a trend that could accelerate the development cycle itself.

The debate has also been sharpened by the resignation of Anthropic researcher Jacob Coxon, who said leading AI companies were taking unacceptable risks. Amodei’s post did not directly address the resignation, but the episode reflects the trust crisis surrounding frontier labs.

Significance and unresolved questions

The proposal’s most concrete contribution is its attempt to turn broad safety rhetoric into an oversight mechanism. External evaluators with meaningful access could reduce the limits of self-assessment, while common standards might prevent companies from being punished competitively for investing in safety.

Yet significant obstacles remain. Coordination among companies could attract antitrust scrutiny, while cooperation with China is constrained by strategic competition and technology controls. Critics also warn that rules shaped by a small group of leading labs could raise barriers for newcomers and become a form of regulatory capture. Others question whether catastrophic AI scenarios deserve priority over existing harms such as bias, labor disruption, and misuse.

“Pacing the frontier” is therefore better understood as a governance experiment than a finished policy. Its credibility will depend on evaluator independence, transparent incident reporting, and whether the commitments apply broadly rather than only to a few prominent companies. Amodei continues to argue that AI can significantly improve human life, but says those benefits require using the time gained to develop the technology more deliberately.

Source: TechCrunch AI