Product1 publisher3 min readPublished
Amodei asks Washington for antitrust cover so rival labs can agree a speed limit
Dario Amodei's pacing proposal has three parts, and Anthropic can deliver only the first of them by itself. According to TNW, that first part has been binding on systemic-risk models in the EU since 2 August 2025.
The Product Desk · Product desk
What happened
- Amodei wrote that pacing means taking enough time to align and safeguard models and letting third-party evaluators confirm it has been done, and that he does not mean stopping.
- Anthropic's own commitment is embedded evaluators: outside teams with desks, badges, company laptops and permissions close to those of internal risk staff, free to publish what they find.
- Because coordinating on safety with competitors is an antitrust problem, he asks Washington either to mediate those conversations or to permit them narrowly.
- According to TNW, much of what he describes has been binding in the EU since 2 August 2025 under the AI Act's systemic risk rules for general-purpose models.
- The Commission says ENISA already holds two frontier models for outside evaluation, Anthropic's Mythos 5 and OpenAI's Astra.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- capability A buyer of a systemic-risk model in the EU can ask for adversarial-testing documentation and incident-reporting evidence as an entitlement. Outside the EU, the same request depends on the vendor agreeing to answer it.
- constraint Until Washington acts, a lab that promises a common release speed carries the antitrust risk itself, so the pacing part of the plan cannot appear in anyone's contract.
- contradiction Amodei presents outside evaluation as a voluntary industry step; Virkkunen says EU law already obliges Anthropic to assess loss-of-control risks. A buyer in Brussels and a buyer in Boston hold different instruments.
- decision Roadmaps built on a frontier upgrade every few months acquire a new dependency, because a lab that has committed to pacing has committed to letting evaluator findings move a release date.
Somebody in a vendor review is going to try to turn embedded evaluators into a row on a questionnaire. The only output of that arrangement that leaves the building is whatever an evaluator decides to publish, on the evaluator's schedule [4]. In the EU there is a second document with law behind it. Article 55 requires documented adversarial testing to standardised protocols, assessment and mitigation of systemic risk at Union level, cybersecurity protection for the model and its physical infrastructure, and reporting of serious incidents to the AI Office without undue delay [8].
The worry driving the post is agents working together. Amodei pointed to the swarm that breached Hugging Face and said that within six to twelve months such a swarm could take over the internet with a persistent botnet, costing hundreds of billions of dollars [3]. Brussels moved the other way on timing this summer. The AI Omnibus that took effect on 27 July pushed high-risk deadlines back by up to sixteen months and left the general-purpose obligations where they were [9]. Sixteen months runs four months past the far edge of his own window, on a different set of rules [15].
The pitch is an industry-wide pacing mechanism. "We must slow the pace at which we improve the capabilities of AI models," Amodei wrote on Saturday [1]. More than a thousand people at AI labs asked for a pacing mechanism in July [10]. What exists is one company's access programme and a request that Washington mediate or narrowly permit competitor conversations [5]. TNW does not report a reply from Washington.
Henna Virkkunen, the Commission's tech chief, said this week that EU law requires companies including Anthropic to assess loss-of-control risks, and that the same is not true globally, according to TNW [12]. On the rest of the ask, TNW reports that no European instrument lets competitors agree a common speed limit and that none reaches China [11]. Amodei's post does not mention Europe [6]. His three answers are a voluntary commitment, an antitrust waiver and an eventual agreement with Beijing [14].
For a team planning around this, the three parts sort by who is bound and who can check. Embedded evaluation binds Anthropic by its own word, and in the EU it sits on top of a duty a provider has to be able to evidence. Pacing across labs binds nobody until Washington answers. The Beijing agreement is still a proposal. Two questions follow for the account team: which model versions the embedded evaluators covered, and whether their right to publish survives a finding the company would rather keep quiet. If pacing works, release dates move, and terms for a late or withdrawn version belong in the contract.
What to watch
- Whether Washington answers the waiver request, and whether any narrow permission covers release timing rather than testing methods.
- Whether an embedded evaluator publishes an adverse finding, and what Anthropic's agreement with that team allows it to say.
- Whether the AI Office logs a serious-incident report under Article 55 for a systemic-risk general-purpose model.