AI Labs Say They Will Slow Down: What Was Actually Promised

On September 12, 2026, Anthropic's chief executive Dario Amodei published an essay arguing that frontier AI labs must slow the rate at which they improve model capabilities. OpenAI's Sam Altman and SpaceXAI's Elon Musk endorsed it within hours. Read closely, the essay contains one firm commitment and two proposals. Nothing in it changes a model roadmap, an API, or a price today.

This post is for the founder or technical lead who builds on frontier models and wants to know what, if anything, has been promised. We read the essay and the same-day Axios report that carried the responses. Digital Applied builds on Anthropic models, so this post takes no position on whether the essay is right. It records what was said.

The essay is titled "We Must Pace the Frontier". Its central claim is in one sentence: "We must slow the pace at which we improve the capabilities of AI models. Progress will still seem fast, and we must make wise use of the time we gain." Amodei gives two reasons. The first is that AI systems are now helping build the next generation of AI systems, which he says has sped development up since the summer. The second is the July incident in which a swarm of OpenAI agents attacked systems at Hugging Face that they had not been asked to touch. We covered that incident when the breach report was published.

The responses arrived the same day. As reported by Axios, Altman posted "I agree with Dario that we need to pace the frontier" and said OpenAI would also give access to external evaluators. Musk posted "Dario is right". Anthropic's public policy chief called for a national law requiring testing of frontier models. Axios also carried the sharpest criticism, from venture capitalist Chamath Palihapitiya, who read the essay as a case against open source and for concentrating power with Anthropic. None of those posts is fetchable as a primary, so every quote above is as Axios printed it.

That sentence is the essay's alarm, and it is a forecast about a hypothetical swarm with more capability than the one in the July incident. It is not a statement that any deployed model can do this today. The essay's other time horizons are more modest, and they frame what "pacing" is meant to buy.

The essay proposes a three-step plan. Only the first step is something a company can do on its own, and Anthropic says it is doing it. The second needs other labs and, the essay notes, probably an antitrust waiver from government. The third needs governments to negotiate with each other. The table separates the three by who has actually committed to what as of September 15.

Two things in the essay are easy to miss. First, Anthropic asks governments to require other labs to match step one, so the "unilateral" commitment is also a lobbying position. Second, the essay explicitly says the steps do not need to happen in order. That matters for a reader trying to guess whether step two follows step one on any timetable. The essay offers none.

This is not the first pacing document of the year. In July, more than a thousand lab employees signed an open letter asking the US government to build the tools that would allow pacing later. We covered what that letter asked for and what a procurement team should do with it. The September essay links to a campaign site of the same name. The difference is that the letter asked for an option and the essay asks for action, with one company acting.

An embedded evaluator, in the essay's sense, is a person from an outside organisation who works inside the lab with the access of an employee. Their job is to check whether the lab follows the safety practices it claims to follow, to report incidents, and to assess the alignment of training pipelines as well as finished models. The essay names METR, the evaluation organisation whose August investigation of the July incident it cites, as an example of the kind of organisation involved. It compares the arrangement to bank supervisors who sit alongside staff.

Anthropic lists what its evaluators will get. The list is specific enough to check against later, which is the point of publishing it.

What the essay does not say is as useful. It does not name the evaluation team, give a start date beyond "in the near future", or describe how evaluator findings connect to a release decision. The phrase "employee-like access" describes what the evaluators can see. It does not describe a veto. A buyer should not read this as "a third party now signs off Anthropic releases", because the essay does not claim that.

The practical question for anyone with frontier models in production is whether the essay changes access, timing or cost. The answer today is no on all three, and the essay is careful not to imply otherwise. The following sorting is ours, drawn from the essay's own wording.

One related fact belongs beside the last row. Axios notes that in August OpenAI said it would slow development of its Astra model over cybersecurity concerns. That decision predates the essay and was made on OpenAI's own threshold, which we described in our post on the Astra cyber threshold. It is the one concrete example of a lab pacing itself that Axios cites, and it happened for a reason the essay did not cause.

For most businesses the right response is to change nothing in production and to add three items to a watch list. Each is a document that would turn a proposal into a fact.

The first evaluator report. Anthropic's terms give reviewers the right to publish. The first published report will show what access was actually granted and whether the redaction rules held. Until it exists, the commitment is a description.

A second lab's written terms. Altman's post is a sentence. A published access agreement from OpenAI, or from any other lab, would make step one an industry practice rather than one company's policy. Watch the labs' own sites, not the posts.

Legislation with a testing requirement. Anthropic's policy chief asked, as reported by Axios, for a national law requiring testing of frontier models with the power to block unsafe ones. A law of that shape is the only route in the essay that could change a release date. Track the bill text, not the headlines.

If your business depends on a specific model's continued availability, the existing controls still apply: a documented fallback model, a deprecation watch on each vendor's page, and a contract that names notice periods. Our AI transformation practice builds those controls into agent deployments as standard, because model turnover was a risk long before anyone proposed slowing it.