Pace is not a pause

Two things happened this month that look like they contradict each other.

In September, Dario Amodei, who runs Anthropic, published an essay called We Must Pace the Frontier. His argument is that the labs should slow the rate at which AI capabilities improve so that safety work can keep up. He is careful about the verb. He is asking for a slower pace, not a halt.

Then, on September 22, Anthropic released Claude Opus 5.5 and OpenAI released GPT-6 Sol and GPT-6 Luna. Anthropic says Opus 5.5 costs less per token than the model it replaces, and Luna is among the cheapest models OpenAI has shipped.

If you only read headlines, it looks like hypocrisy or noise. If you run a business on these tools, it is useful information. For a structure that holds up whichever way it goes, see The AI-first company, for owners.

What the essay actually asks for

The essay is specific. Anthropic commits to bringing in outside evaluators, such as METR, with access to its offices, tools and permissions close to what employees have, and the right to publish what they find. It asks other frontier companies to do the same. It asks democratic governments to make it legal for the companies to coordinate on safety and to regulate through transparency and audits. It says any slowdown must not give up America’s lead over Chinese AI capabilities. And it warns that, if capability growth goes unchecked, misaligned AI systems could do serious damage to internet infrastructure within six to twelve months. That last point is his forecast, not something that has happened.

None of that says “stop shipping.” It says ship with more eyes on the work.

What an owner should take from both

Capability will keep arriving. Even the most cautious leader in the industry is asking for a slower climb, not a flat line. Plan for models that get better and cheaper every few months, because that is what every lab is still doing.

The rules around the models will tighten. Outside evaluators, audits and transparency requirements are coming in some form. Expect more disclosure about what a model can do, and more limits on what it is allowed to do inside regulated work.

Your systems should not care which model is inside them. This is the practical lesson. If a new model ships next month, or an old one is restricted, a well-built business swaps the part and keeps running.

A swap test you can run this week

Take the three workflows that matter most to your business and ask, for each one:

1. Which model does it use today, and where is that written down?
2. If that model disappeared tomorrow, how long would it take to move the workflow to another one: an hour, a day, a month?
3. What would break: the prompts, the tools, the output format, or the checks you rely on?

Anything that takes longer than a day to move is a risk. Fix those first. The goal is not to predict which lab wins. It is to make the answer not matter.