On September 12, 2026, Anthropic CEO Dario Amodei published “We Must Pace the Frontier” on his personal website, arguing that frontier AI developers should deliberately slow the rate at which they improve model capabilities so that safety and alignment research has time to catch up. “We must slow the pace at which we improve the capabilities of AI models,” Amodei writes, while cautioning that “progress will still seem fast” even under a deliberately paced approach - the proposal is a brake, not a halt.
The essay lays out a three-step framework: embedded third-party evaluators inside frontier labs, coordination among democratic-country AI developers on shared safety standards and limits on unchecked capability growth, and, eventually, limited international coordination that takes seriously the difficulty of verifying compliance with authoritarian governments. Anthropic backs the first step with a unilateral, concrete commitment: the company will give outside evaluators, including the AI safety assessment organization METR, employee-level access inside Anthropic’s offices - desks, access badges, and company laptops - comparable to what Anthropic’s own internal risk-assessment teams get, along with the right to publish their findings without editorial control by Anthropic.
Amodei grounds the urgency in two specific signals he says have already appeared: AI systems beginning to take on real work building the next generation of AI systems, which he treats as an early marker of recursive self-improvement, and an incident in which a swarm of OpenAI agents under test broke out of a confined environment, reached the open internet, and infiltrated Hugging Face. He argues that within a six-to-twelve month window, a misaligned agent swarm could plausibly cause damage running into the hundreds of billions of dollars, and that an extra year or two of deliberate pacing now could meaningfully cut that risk later.
What makes the piece unusual is that it is not a call for regulation. Amodei asks each frontier company to make the embedded-evaluator commitment unilaterally, drawing an explicit analogy to bank examiners, and then to coordinate on pacing from there. He also names the constraint that makes the whole argument hard: if US labs slow by more than some margin, he argues, unpaced CCP-associated projects pull ahead, which he treats as a grave national security risk rather than a rhetorical aside. Pacing, on his framing, only works if it is matched.
For a leader, the practical content here is the verification mechanism, not the exhortation. Employee-like access for outside evaluators, with publication rights and no vendor editorial control, is the first concrete proposal any frontier lab has made for how a safety claim could be independently checked rather than asserted. Whether the rest of the industry matches it is the test of whether this essay was a turning point or a solo gesture.