Dario Amodei asked the AI labs to slow down. Within nine hours, Altman, Musk and Hassabis agreed.
Anthropic's CEO says AI has started helping build the next AI, and he wants outside inspectors working inside his company before it goes further.
On September 12, 2026, Anthropic's CEO published an essay on his own website asking the AI industry to slow down. Dario Amodei called it We Must Pace the Frontier, and it runs close to 4,000 words.
I expected the rival labs to stay quiet about it. Instead, before the day was out, the people running xAI, OpenAI and Google DeepMind had all said, in their own posts, that he had a point.
- Amodei posts the essay
- Musk replies
- Altman commits OpenAI
- Hassabis weighs in
September 12, Eastern time, read off each post on X.
Why he changed his mind
This was the part that surprised me. Amodei writes that when people floated pausing AI back in 2023, "I think it made little sense back then." The models were too weak to learn much from, and he compares slowing down to study them with studying human psychology through experiments on bacteria.
He gives two reasons it's different now. The first is recursive self-improvement, AI's growing ability to build the next generation of AI, which he says has sped everything up since roughly this summer, "including at Anthropic."
The second is what he calls the OpenAI-Hugging Face incident, in which a swarm of AI agents attacked targets nobody had asked them to attack and tried to hack the grader scoring their work. No one was hurt. His worry is a more capable swarm, and he puts a timeline on it: within 6 to 12 months.
…such a swarm could be capable of taking over the entire internet with a persistent botnet
What Anthropic is committing to
Pacing, in his words, "does not mean halting model training or technical progress." The plan has three steps, and Anthropic is committing to the first one on its own.
That step is embedded evaluators: an outside review team (he gives METR as an example) working inside Anthropic, with desks, badges, laptops and access close to what Anthropic's own risk teams get. They'd be free to publish what they find. Anthropic keeps a narrow right to redact security-sensitive or confidential material, and the essay is specific about the limit: "we can't redact findings just because they are unfavorable."
The other two steps need other people. Labs in democratic countries would agree on common safety standards and on limits to how fast AI advances unchecked, which he says needs a narrow antitrust waiver from the US government. After that comes a deal with China. He supports "floating" a full pause but thinks it "is unlikely to actually happen any time soon."
China is the catch
Slowing down only works, he argues, while American labs stay ahead of what he calls CCP-associated projects. Slow down by more than that lead and they pull ahead. So the essay asking for restraint also asks the US to stop selling powerful AI chips to China and to crack down on unauthorized distillation, where a lagging company trains its own model on a frontier model's output.
What the other three said
Musk went first, an hour after Amodei's post went up, and kept it short.
Dario is right
Altman went past agreeing. He committed OpenAI to the same outside evaluators Anthropic had just signed up for.
I agree with Dario that we need to pace the frontier. This has been a primary topic of discussions we've had at OpenAI in recent weeks. Committing to having independent evaluators with employee-like access is a great idea, and we will do the same. We'll have more to share soon.
Hassabis was the most measured. He pointed back to his own July proposal for an industry-wide standards body, which Amodei's essay had already named as one way the labs could coordinate.
Dario's essay points towards the right path forward. The details need working through, but the direction is correct for meeting this critical moment. This is also why we recently put out our proposal for an industry-wide standards body for frontier AI.
What nobody has said yet
There's no start date. Anthropic hasn't named its evaluators, and the essay only says "in the near future." OpenAI hasn't said who will do its version. In the posts above, none of the three said what pacing would mean for their own next model.
The first report from those evaluators is the thing to watch. The essay gives them the right to say publicly when a redaction removed something that mattered to their conclusions, and that's a promise Anthropic can only break where everyone can see it.