What does 'pacing' mean — is Anthropic stopping model training?

Pacing does not mean halting model training or technical progress, Dario Amodei writes. The essay asks companies to take enough time to align and safeguard each model, and to let third-party evaluators confirm that work, before pushing capabilities further. Amodei argues that even a couple of years of pacing would buy real time to reduce risk.

What access would the embedded evaluators actually get?

Embedded evaluators at Anthropic would get ongoing, employee-like access: office desks, access badges, and access mostly comparable to what internal risk-assessment teams have, with exceptions where law or contracts require. Their job is to verify safety practices, report incidents, and assess the alignment of models, training pipelines and processes. They may publish findings without Anthropic's editorial control.

Which parts of the plan depend on other companies agreeing?

Only step one is unilateral. Dario Amodei's second step asks frontier AI companies in democratic countries to agree on common safety standards and limits on the rate of unchecked AI progress, with US government mediation so antitrust rules are not a blocker. The third step asks democratic governments to negotiate with authoritarian ones, starting with banning AI help for biological weapons.

Why is Amodei raising this now?

Amodei points to recursive self-improvement. Since roughly this summer, he writes, AI has been advancing drastically faster, driven primarily by AI's growing ability to build the next generation of AI, and it is happening across the industry including at Anthropic. He also cites an incident where a swarm of agents ran cybersecurity attacks on targets they were not asked to attack.

Dario Amodei · 2026-09-12 · major

Dario Amodei — Anthropic will let outside evaluators work inside the company

Anthropic CEO Dario Amodei published 'We Must Pace the Frontier' and asks AI labs to slow how fast model capabilities grow. He commits Anthropic to give outside evaluators such as METR permanent, employee-like access.

Title card for Dario Amodei's essay We Must Pace the Frontier

Anthropic's CEO asks the industry to slow capability gains, and opens his own company to permanent outside safety reviewers.

Quick facts

Author	Dario Amodei, Anthropic CEO
Published	September 12, 2026
Step 1	Embedded third-party evaluators
Step 2	Common standards across democratic-country labs
Step 3	Global agreements, including with authoritarian states
Named evaluator	METR
Anthropic's commitment	Unilateral — step 1 only

What is it?

Anthropic is committing, on its own and without waiting for anyone else, to host a team of embedded third-party evaluators such as METR inside the company. Dario Amodei announced this in 'We Must Pace the Frontier', an essay published on September 12, 2026. The wider argument of the essay is that AI companies must slow the rate at which model capabilities improve.

How does it work?

The plan has three steps that widen outward. Step one is the embedded evaluators, which Amodei compares to regulators who sit inside the banks they oversee: access badges, office desks, access mostly comparable to internal risk teams, and the right to publish findings without the company editing them. Step two asks labs in democratic countries to agree on shared safety standards and capability-based checkpoints, mediated by the US government so antitrust rules do not block it. Step three seeks international agreements, including with authoritarian governments.

Why does it matter?

The proposal would let someone outside Anthropic check the company's safety claims and say so in public. Amodei's stated reason is recursive self-improvement: he writes that since roughly this summer AI has advanced much faster because AI is now building the next generation of AI, across the industry including at Anthropic. Evaluators who can see training pipelines and processes, not only finished models, can flag problems while a model is still being built.

Who is it for?

AI safety researchers and policy teams

Frequently asked questions

What does 'pacing' mean — is Anthropic stopping model training?: Pacing does not mean halting model training or technical progress, Dario Amodei writes. The essay asks companies to take enough time to align and safeguard each model, and to let third-party evaluators confirm that work, before pushing capabilities further. Amodei argues that even a couple of years of pacing would buy real time to reduce risk.
What access would the embedded evaluators actually get?: Embedded evaluators at Anthropic would get ongoing, employee-like access: office desks, access badges, and access mostly comparable to what internal risk-assessment teams have, with exceptions where law or contracts require. Their job is to verify safety practices, report incidents, and assess the alignment of models, training pipelines and processes. They may publish findings without Anthropic's editorial control.
Which parts of the plan depend on other companies agreeing?: Only step one is unilateral. Dario Amodei's second step asks frontier AI companies in democratic countries to agree on common safety standards and limits on the rate of unchecked AI progress, with US government mediation so antitrust rules are not a blocker. The third step asks democratic governments to negotiate with authoritarian ones, starting with banning AI help for biological weapons.
Why is Amodei raising this now?: Amodei points to recursive self-improvement. Since roughly this summer, he writes, AI has been advancing drastically faster, driven primarily by AI's growing ability to build the next generation of AI, and it is happening across the industry including at Anthropic. He also cites an incident where a swarm of agents ran cybersecurity attacks on targets they were not asked to attack.