What it means that Anthropic and OpenAI agree to slow down the pace of their AI models — September 13, 2026
· CompaniesAutomation
Flash Edition of the Radar: Dario Amodei publishes "We Must Pace the Frontier" and calls for a slowdown in model improvement; Anthropic accepts embedded external evaluators with employee access and no editorial veto, and OpenAI says it will do the same. What clause to demand from your provider and how to lock down your agents today.
Flash Edition. The CEO of Anthropic is calling for a deliberate slowdown in the pace of model improvement, and OpenAI says it will do the same. A supplement to this morning's Radar and a follow-up to the September 11 flash.
What Happened
On September 12, Dario Amodei published the essay "We Must Pace the Frontier": "we must reduce the rate at which we improve the capabilities of AI models." He argues that models are already helping to build the next generation and that, following July's swarm of agents, in 6-12 months a larger one could "take over the entire internet with a persistent botnet," causing "hundreds of billions of dollars" in damages.
The plan has three steps. First, embedded external evaluators —like METR— with employee access: a desk, badge, laptop, and the right to publish their findings without editorial control from the company; Anthropic is committing to this unilaterally. Second, common standards among democratic labs, with Washington mediating against antitrust concerns. Third, international agreements, including China. Sam Altman responded that OpenAI agrees and will implement the same, with details to follow (TechCrunch).
Why It Matters
This is not a philosophical debate: it is the first concrete commitment that your AI provider will allow a third party to come in and see how they work and report what they see. Two consequences. One: "the model will fix it next year" is no longer a plan; if the frontier is paced, what you have poorly built will stay poorly built. The other, more important one: the role Anthropic just accepted for itself —an auditor with real access and no veto over the report— is what you should demand in your contract.
For Your Company
Three things, none of which cost money. One: in your next renewal, ask in writing for two clauses that almost no one asks for today: notification of model incidents within a fixed timeframe, and a third-party audit of the workflow that affects you without the provider being able to veto the report. If Anthropic grants this to METR, your integrator has no excuse. Two: if you have agents in production, apply the lock from the September 11 flash: logs are written outside the agent, in append-only storage with credentials the agent does not possess, and the alarm triggers on the new privilege, not on the accomplished damage. Three: cut back the effective permissions of your integrations this week. Those on steps 2 or 3 of the AI First ladder win with this slowdown: room to consolidate before the next wave.
Frequently Asked Questions
Does this mean AI will advance more slowly and I can relax?
No. Amodei says "progress will still feel fast." The distance between what the model can do and what your company knows how to do with it is not narrowed by the lab: it is narrowed by you.
Is it a binding commitment?
No. It is a personal essay, a unilateral commitment from Anthropic, and an endorsement from OpenAI without published details. There is no law, no designated auditor, and no penalty: it is a signal, not a guarantee.
What if my provider is not Anthropic or OpenAI?
All the more reason. The commitment from the two giants creates a citable market standard: ask your provider which external evaluator they submit to and what they publish when something goes wrong. The answer, or lack thereof, is already information.