Dario Amodei's essay "We Must Pace the Frontier" asks frontier labs to slow capability gains enough for alignment, interpretability, and testing to catch up. The step he says Anthropic will take alone is embedded third-party evaluators. The same month, Anthropic and OpenAI still shipped cheaper new models.
Published · 2 sources
Likely. That Amodei made this argument, including the embedded-evaluator commitment, is his own essay, so the text is not in doubt. The confidence stays at "likely" because it is one executive's case for what the industry should do, not a demonstrated change in release practice. A published evaluator agreement, or a model Anthropic says it held back under this policy, would be what would move this entry.
The practical test of a slowdown is what shipped afterwards. Amodei argues for more time before the next capability jump, and for outside reviewers who can see training and deployment the way employees do. Ten days later Anthropic released Opus 5.5 and OpenAI released GPT-6 Sol and Luna. Those are real products you can use. They are also evidence that "pace" did not mean "stop."
If you depend on these models at work, the essay does not change your API. It tells you the chief executive of one lab thinks recursive self-improvement — models helping build the next models — is already underway, and that he wants that loop checked by people who do not work for the lab.
Dario Amodei
darioamodei.com
Primary source
Open original (opens in a new tab)The Straits Times
straitstimes.com
Secondary
Open original (opens in a new tab)Learn the idea behind this story
The essay is about whether builders can show their work. The lab is the small version of that: write the task so someone else can see what you asked for.
Open the prompt lab