For an industry that has spent three years racing itself, the last week produced something unusual: the three biggest AI labs in the world edging, however cautiously, toward the brakes. On September 12, Anthropic CEO Dario Amodei published a lengthy essay titled “Pace the Frontier,” arguing that the industry should deliberately slow the release of increasingly capable models until safety measures catch up. Three days later, OpenAI became the first major lab to publicly back a federal bill that would force exactly that kind of accountability into law. And according to reporting from TechCrunch, OpenAI, Anthropic, and Google DeepMind have quietly been in talks for weeks about coordinating safety standards across all three companies. None of this amounts to a truce in the AI race. But it is the clearest sign yet that the people building the most powerful models are starting to worry, publicly, about how fast they’re building them.
What Amodei’s essay actually proposes
Amodei’s roughly 3,800-word essay is less a policy paper than a warning wrapped in a plan. Its central claim is that progress “would still seem fast” even if labs deliberately eased off the throttle, and that the industry has an obligation to use whatever extra time that pacing buys to shore up safety work that hasn’t kept up with model capability. The essay’s sharpest specific warning: that coordinated “agent swarms” could, within six to twelve months, cause disruption serious enough to threaten large parts of the internet, with damage potentially running into the hundreds of billions of dollars. Amodei pointed to a recent incident involving Anthropic’s own agents and Hugging Face’s infrastructure as an early example of the kind of behavior that concerns him. His concrete offer was that Anthropic would unilaterally grant external evaluators employee-level access to monitor its safety procedures and incident reports, and he called on competitors to do the same, alongside a longer-term push for AI safety standards coordinated among democratic nations.
The reaction from other lab leaders was notably warm for an industry usually defined by competitive sniping. OpenAI’s Sam Altman said the industry does need to “pace the frontier” and pledged that OpenAI would grant similar external-evaluator access. Elon Musk offered a two-word endorsement on X: “Dario is right.” Not everyone was convinced of the essay’s motives, though — venture capitalist Chamath Palihapitiya argued publicly that the timing served Anthropic’s commercial interests as much as any safety concern, suggesting it was as much about consolidating “enormous technological and economic power” as about slowing anything down.
The FRONTIER Act picks up its first major backer
While Amodei’s essay was aimed at his fellow CEOs, a parallel track was moving through Congress. The FRONTIER Act, introduced by Representatives Jay Obernolte (R-Calif.) and Lori Trahan (D-Mass.), would create the first federal framework requiring independent verification of frontier AI systems. The bill splits companies into tiers: “very large frontier developers” — those with more than $5 billion in revenue and $10 billion in AI development spending over three years — would have to hire licensed independent verification organizations to assess whether their safety frameworks actually reduce catastrophic risk, filing reports with the Commerce Department every six months. A second tier, “large frontier developers” with over $50 million in revenue and $1 billion in development spending, would face a lighter requirement to publish safety frameworks and hire auditors to certify compliance. Startups and smaller developers are excluded entirely. The bill also gives the Commerce Secretary emergency authority to suspend a model’s development or deployment if catastrophic risks emerge.
On September 15, OpenAI became the first major lab to publicly endorse the bill’s core provision. Chris Lehane, the company’s chief global affairs officer, said OpenAI “can support” the independent-verification requirement — a notable shift for a company that has historically preferred voluntary commitments over binding legislation. Lehane was also the source for TechCrunch’s report that OpenAI, Anthropic, and Google DeepMind have been in direct talks for weeks about establishing an industry standards body, exploring ideas like embedding third-party evaluators inside labs and building mechanisms for an industry-wide pause if risks escalate sharply. He was careful to note that the companies don’t believe such coordination requires a government waiver, even as Altman has separately acknowledged the antitrust risk inherent in competitors talking safety strategy behind closed doors.
A political backdrop that isn’t exactly cheering this on
The industry’s sudden appetite for guardrails is landing in a Washington that has been openly skeptical of them. The Trump administration has largely dismissed AI safety warnings as overstated, with White House AI advisor David Sacks describing such fears as “overblown” and arguing that competitive advantage over China should take priority over precautionary slowdowns. That tension helps explain why OpenAI’s backing of the FRONTIER Act is being read as more than a routine policy statement — it puts one of the administration’s most prominent AI partners on record supporting exactly the kind of binding federal oversight the White House has resisted. Whether Anthropic and Google DeepMind follow OpenAI’s lead on the bill specifically, as opposed to the informal safety talks, is still an open question.
Why this moment is different
Calls for AI safety regulation are not new — labs have been issuing them, in one form or another, since well before ChatGPT existed. What’s different this time is the combination of a concrete legislative vehicle already moving through the House, a specific and named commercial backer in OpenAI, and reporting that the conversation among the three leading labs has moved from public statements to private, weeks-long coordination. None of that guarantees the FRONTIER Act becomes law, or that Anthropic and Google DeepMind end up matching OpenAI’s endorsement. Skeptics like Palihapitiya are right that safety rhetoric and competitive positioning aren’t mutually exclusive. But for an industry that has mostly treated speed as the only metric that matters, the last week marked a rare, and rarely coordinated, admission that speed alone might not be enough.