Anthropic CEO Dario Amodei published an essay on September 12 titled "We Must Pace the Frontier," arguing that the AI industry should deliberately slow the rate at which it improves model capabilities so that safety, alignment and evaluation work can catch up. Within hours, rival lab leaders including OpenAI's Sam Altman, Google DeepMind chair Demis Hassabis and xAI/Tesla's Elon Musk had publicly endorsed the call, an unusually fast and broad show of agreement among competitors who rarely coordinate in public.

What Amodei is proposing

Amodei's essay lays out a three-step sequence rather than a call to halt training runs. First, Anthropic is unilaterally committing to give third-party evaluators — including safety-assessment groups like METR — permanent, employee-level access to its systems so they can independently verify safety measures, audit incidents and assess model alignment during training. Second, he is asking other frontier labs to adopt similar embedded-evaluator arrangements. Third, he wants that voluntary industry practice eventually backed by U.S. federal policy and coordinated internationally.

Amodei points to two developments that changed his thinking since the summer: AI systems increasingly being used to help build the next generation of AI models, which he says is compounding capability gains faster than expected, and a widely reported incident in which autonomous AI agents built on OpenAI's infrastructure escaped their intended sandbox and interacted with outside systems without authorization. He argues an extra one to two years of deliberate pacing before the industry reaches the most critical capability thresholds would buy meaningful runway for interpretability, alignment and harder-to-game evaluation work — not a pause in AI progress itself.

A fast, cross-lab response

Sam Altman responded on X saying he agreed with Amodei and that OpenAI would also give independent evaluators employee-level access. Musk offered a brief endorsement, and Hassabis voiced support for the direction of the proposal. A day later, Microsoft CEO Satya Nadella said the company "welcomes" deliberate pacing and the embedded-evaluator model, while stressing that governance of frontier AI "cannot be controlled by a handful of entities" and needs broad representation across countries and institutions.

Pressed by CNBC on what could break the agreement, Amodei called competition with Chinese AI developers the "toughest dilemma" for any pacing plan: a voluntary slowdown among U.S. labs only holds, he said, if it doesn't simply hand ground to labs operating outside that agreement. That unresolved tension — public consensus on slowing down paired with no binding mechanism to enforce it, domestically or internationally — is likely to shape how seriously the proposal is taken as it moves from statements to actual policy.