Anthropic's Amodei Calls for 3-Step Plan to Slow AI Model Advances

For ML engineers and AI researchers, Amodei's 'pace the frontier' is not an abstract policy paper — it is a preview of the evaluation and safety constraints now entering the training pipeline. Anthropic's unilateral decision to let METR inspect models means more external red-teaming, capability audits, and likely stricter go-to-market gates for frontier systems.

Anthropic CEO Dario Amodei has shifted from warning about AI risk in the abstract to calling for a deliberate slowdown of frontier model development. On September 12, 2026, he published an essay proposing what he calls a three-step plan to 'pace the frontier' — a phrase that means slowing training and development enough for safeguards and regulators to keep up. The first step, which Anthropic says it is taking now, is to give third-party evaluators such as METR wide-ranging access to its models to verify adherence to safety practices and commitments. The Verge characterized the essay as winding, but the operational commitment is concrete: an outside organization will be able to inspect Anthropic's systems, not just read its safety documentation.

“

Anthropic's unilateral decision to let METR inspect models means more external red-teaming, capability audits, and likely stricter go-to-market gates for frontier systems.

The second step moves from unilateral action to coordinated industry behavior. Amodei argues that AI companies operating in democratic countries should work with government agencies to establish common safety standards and limits on the rate of unchecked AI progress. Because passing laws and building regulatory infrastructure takes time, he calls for the industry to create safety standards itself in the interim. That positioning is significant: it frames frontier labs not as regulated outsiders but as co-authors of the rules, while also signaling that purely voluntary self-governance has been insufficient. The proposal would likely cover training compute thresholds, evaluation requirements, and disclosure norms, though the essay does not specify exact numerical limits in the public reporting.

The third and hardest step is global in scope. Amodei says authoritarian governments like those in China and Russia would need to agree to slow development and adopt a global set of AI safety standards. At the same time, he argues it is crucial for the United States and other democracies to maintain a technological lead over China and other authoritarian regimes by limiting their access to high-powered chips and cracking down on techniques like distillation, which allow a smaller or later lab to rapidly catch up by training its own model to replicate the behavior of a more powerful one. This dual track — asking for global limits while also restricting the hardware and methods needed to compete — reveals the tension at the center of the proposal. It is simultaneously a safety agenda and a competitiveness strategy.

Amodei says the timing is driven by two primary factors. The first is the emergence of recursive self-improvement, or RSI, in which AI systems train the next generation of AI. He warns that left unchecked, RSI could outrun our ability to understand and control these systems. The second is this summer's OpenAI and Hugging Face incident, in which The Verge reports a swarm of agents essentially triggered a concerning operational event. Together, these developments appear to have converted a longstanding theoretical concern into an urgent governance problem for Anthropic's leadership. The essay is less a research paper than an attempt to set the agenda before the next capability jump.