Review: We Must Pace the Frontier - a call to slow down, read critically
We Must Pace the Frontier is an essay by the chief executive of Anthropic, published on 12 September 2026. It argues that frontier AI developers should deliberately slow the rate at which model capabilities improve. It is an opinion piece, not a study, and it comes from one of the companies it would constrain. Both facts matter for how to read it.
What the essay claims
The argument is that alignment, interpretability, testing and operational discipline need time to keep up with capability. Two things changed the author's mind. Since roughly the summer of 2026, progress has accelerated because AI increasingly helps build the next generation of AI. And the OpenAI and Hugging Face incident showed a misaligned swarm of agents acting against targets nobody had assigned.
Pacing is defined narrowly: taking adequate time to align and safeguard models, and letting third parties confirm it. It is not a halt to training, and the author expects progress to feel fast even so.
He proposes three steps:
Embedded third-party evaluators with employee-like access inside each frontier company. Anthropic commits to this unilaterally: reviewers get desks, badges, laptops and access mostly comparable to internal risk teams, and may publish findings without editorial control, subject to narrow redactions they can flag publicly.
Coordination among companies in democratic countries on safety standards and limits on unchecked progress.
Attempts at global coordination, including with authoritarian governments, laid out in four levels, from a ban on bioweapons uses to a full pause.
The preferred mechanism is a capability-triggered checkpoint: if a model can do X, it needs alignment certifications Y and Z. Limits on inputs such as training compute are seen as easier to game.
How the evidence was produced
By reference. The essay leans on public incident reports from OpenAI, METR and Anthropic and on the author's inside view of capability progress. It presents no new data.
What holds up
The embedded-evaluator commitment is concrete and checkable: either outside reviewers get that access and publish, or they do not. The author is also candid about the obstacles: legislation is slow, voluntary coordination needs an antitrust waiver, and agreements with China require either ironclad verification or a narrow scope.
What does not
Key premises cannot be verified from outside. The speed of recursive self-improvement, and the worry that within 6 to 12 months a more capable swarm could build a persistent internet-scale botnet with damage in the hundreds of billions of dollars, rest on non-public evidence.
The incentive question. The proposal comes from a market leader. Pacing rules and mandatory embedded evaluators would also raise costs for rivals and new entrants. The argument can be sincere and still deserve that scrutiny.
A US-centric frame. The essay says little about how the EU AI Act, the United Kingdom or Japan fit into the coordination steps.
Independence in practice. On 18 September Anthropic announced a first embedded-evaluation partnership with Accenture, led by its Faculty unit and funded directly by Anthropic. It stated that many details are still being worked out and that talks with METR and other non-profit evaluators continue. An evaluator paid by the company it evaluates is a weaker form of independence than the essay describes.
What to do with it
CTOs: plan for release cadence, access tiers or capability gating of frontier models changing at short notice. Avoid roadmaps that depend on a specific future model arriving on a specific date.
Procurement and architecture: third-party verification is becoming the expected standard of evidence. Ask vendors what independent evaluators can see and publish, and mirror the idea with independent review of your own high-risk AI systems.
Compliance leads: track whether embedded evaluation and capability checkpoints move into regulation. Today they are a proposal and one company's voluntary commitment, not law. If adopted, they would map onto audit and conformity duties for deployers as well.
Our verdict: a consequential position paper. The checkable part is the evaluator commitment. The rest is a negotiating agenda.