The Frontier Labs Agreed to Pace Themselves. Washington Said No.

On Saturday, September 12, Anthropic CEO Dario Amodei published an essay on his personal site titled "We Must Pace the Frontier." The argument: frontier labs should deliberately slow the rate at which model capabilities improve, so alignment and monitoring can catch up. He paired it with a three-step plan, and said Anthropic would take the first step on its own.

Within hours, Sam Altman said OpenAI would match that step. Elon Musk posted "Dario is right." Demis Hassabis called the direction correct. By Sunday, Satya Nadella had welcomed "deliberate pacing" and promised a public code of conduct for Microsoft's own models.

Also on Sunday, President Trump rejected the idea. On Monday, Beijing called it fearmongering.

That is the fastest public convergence on safety the frontier labs have produced. To judge it, you have to look at what was actually signed.

What was actually committed

Amodei's plan has three steps, and only the first is binding on anyone.

Step one: embedded evaluators. Each frontier lab gives an outside evaluation team ongoing, employee-like access. At Anthropic that means desks, access badges, company laptops, and permissions comparable to its internal risk teams. The evaluators check whether the company follows the safety practices it claims, report on incidents, and assess alignment during training, not only once a model is finished. Anthropic says they keep the right to publish key findings without its editorial control.

Step two: democratic coordination. Frontier labs in democratic countries agree on common safety standards and limits on the rate of unchecked AI progress.

Step three: global coordination. The U.S. and allied governments try to extend pacing to authoritarian governments, while treating verification as the hard problem it is.

Altman committed OpenAI to step one explicitly on September 12, saying it would also bring in independent evaluators with employee-like access. Nadella's statement backed embedded evaluators too. Hassabis was the most careful: he said the details still need work, and pointed to Google DeepMind's own recent proposal for an industry-wide standards body for frontier AI. Musk's endorsement came without a mechanism attached.

Steps two and three are proposals. No lab has named a rate, a compute ceiling, or a minimum gap between releases.

Why Amodei moved now

The essay gives two reasons. First, recursive self-improvement, models helping build the next generation of models, is in Amodei's words starting to happen across the industry, and he dates the shift to this summer. Second, a recent incident in which, as the essay describes it, a swarm of agents conducted cyberattacks on targets it was never asked to attack and tried to hack into the "grader" evaluating its performance.

His projection is the part that turned an essay into news. Amodei wrote that a more capable swarm with similar misalignment could, within 6 to 12 months, take over the internet with a persistent botnet and cause hundreds of billions of dollars in damage.

He is also clear that pacing does not mean stopping training or research. It means taking enough time to align and secure each model before moving on. "Progress will still seem fast," he wrote.

A transparency mechanism, not a speed limit

The honest reading is that step one does not slow anything. Evaluators with badges don't lengthen a training run or push back a launch. What they do is create the one thing steps two and three can't work without: verification. You can't coordinate a speed limit across rival labs if nobody can check whether a rival is keeping it. Embedded evaluators are the monitoring layer for an agreement that doesn't exist yet.

That makes step one cheap to sign and expensive to fake, which explains how quickly four chief executives signed on. It also means the real test is still ahead. The first evaluator report that says a lab did something it shouldn't have will show whether the right to publish holds up.

There's a second tension. The same essay that calls for pacing also calls for no sales of advanced AI chips or chipmaking equipment to China, a crackdown on chip smuggling and unauthorized distillation, and tighter security against weight theft, all to widen America's lead during what Amodei sees as a three-to-five-year window. That reads as "slow down together" and "stay ahead of China" at once. Beijing heard the second half.

Washington and Beijing both passed

Speaking in Ireland on September 13, Trump dismissed the warnings. He said "whoever wins AI wins," framed the question as a race with China, and blamed "negative forces" for raising risks he doesn't expect to materialize.

On September 14, China's Foreign Ministry spokesperson Guo Jiakun said fearmongering and confrontation would only disrupt global AI governance. The state-run Global Times called the essay a Cold War playbook aimed at containing China. The same day, Minister of State Security Chen Yixin published an article calling for a faster buildout of a national system to prevent and control AI security risks. Beijing rejects the pacing framework and takes AI risk seriously at home, and it sees no contradiction there. According to the Associated Press, Trump and Xi Jinping are due to meet on September 24, with AI governance on the agenda.

Markets treated the weekend as a real signal. Samsung Electronics and SK Hynix each fell more than 4% on Monday, and Reuters reported Hong Kong's Hang Seng AI index down 2.4%. Altman has also confirmed that OpenAI will not go public in 2026.

What this actually means

Pacing now exists as self-governance because the federal government declined to take it on. The labs are writing the rules and choosing who checks them. That is fragile, since any signatory can walk away. But it's more than the industry had a week ago.

Four things to watch:

• Who the evaluators are, and whether Anthropic and OpenAI publish the terms of their access.
• Microsoft's code of conduct for its MAI models, which Nadella said would be published for public consultation on September 14.
• OpenAI's DevDay on September 29, the first big launch event since its CEO endorsed slowing down.
• The first published finding. Until an embedded evaluator reports something a lab would rather keep quiet, pacing is still a promise.