OpenAI vs Anthropic: The Antitrust Waiver Debate (2026)

Dario Amodei asked for a narrow antitrust waiver to make collective AI pacing lawful. Over the weekend, the answer was no — and the structural logic of that refusal clarifies exactly what the request was for.

What Happened

On Saturday, Anthropic CEO Dario Amodei published an essay calling for what he termed “pacing the frontier” — a structured framework under which the leading AI laboratories would coordinate on safety evaluations before shipping their most powerful models. The ask to government was specific and deliberately narrow: “For antitrust reasons, it’s helpful for the US government to mediate or at least enable these discussions — they don’t need to participate, but do need to issue a narrow waiver for certain kinds of safety conversations.” Not regulation. Not supervision. A waiver for a defined category of inter-lab conversation.

Over the weekend, the answer arrived. David Sacks — co-chair of the President’s Council of Advisors on Science and Technology (PCAST), and formerly the White House AI and crypto czar — posted a direct reply on X. He did not deliberate. He granted permission and withheld the accommodation simultaneously: “Dario has written that we need to ‘pace the frontier,’ and Sam has agreed. People may be surprised by my response: go ahead. You guys are the frontier. By any reasonable metric — market share, revenue growth, model capability — the two of you have a duopoly on frontier intelligence.” He then accused the labs of “pretending antitrust law has to be suspended so you can form a cartel.” Sacks speaks as one voice in the administration’s orbit; he is not its formal spokesperson, and no policy has issued from any agency.

Also over the weekend, in a Q&A circulated by @DeItaone and @atrupar — sourced from tape, not from WhiteHouse.gov, and not a primary government record — the president was asked whether AI should slow down or face more regulation. Per that tape, he said the United States leads China in AI and intends to keep it that way, that “whoever wins AI wins,” and that “we could put guardrails, but some voices are being overly negative.” Separately on Sunday, Meta chief AI officer Alexandr Wang stated that alignment is “fundamental” to personal superintelligence and that “alignment can be the gating factor for scaling as we get closer to the frontier” — a statement of principle, explicitly not an endorsement of Amodei’s plan, and not an evaluator commitment. Meta has pledged nothing.

The Structural Read

The Permission Layer framework — which maps how regulatory architecture determines which AI capabilities can ship, and at what speed — is the right lens here. What this weekend clarified is that the permission cost of each tier in Amodei’s plan was always different in kind, not just in degree. The unilateral tier crossed from one firm to two within hours because its permission cost was zero: each company can decide, without asking anyone, what it will not ship. The global tier requires sovereign coordination, which is its own long arc. The industry tier’s permission cost was a narrow antitrust waiver — a specific, bounded legal accommodation — and that accommodation was denied.

Sacks’s central argument deserves to be stated at its strongest rather than waved away, because on its own terms it is correct. If you believe your next model is too dangerous to ship, you need no one’s permission not to ship it. Anthropic demonstrated this directly — the restraint Amodei describes unilaterally is not hypothetical; it has been practiced. Which makes the logic of the waiver request legible: unilateral restraint can do many things, but it cannot do one specific thing. It cannot stop a rival from shipping while you hold back. And that is precisely the outcome competition law exists to prevent labs from agreeing to collectively. The waiver was an attempt to make collective restraint lawful. Refusing it does not abolish pacing; it confines pacing to the unilateral tier, unprotected against anyone who declines to participate — which is the structural vulnerability the essay was written to escape.

The second structural argument Sacks makes is about market position rather than bad faith, which makes it more durable. He grounds the regulatory-capture charge in structure: OpenAI and Anthropic constitute, by his characterisation, a duopoly on frontier intelligence. A rule that only the largest players can comfortably satisfy entrenches those players. Notice the architecture of that argument — it does not require anyone to be lying about risk. Both things can simultaneously be true: the danger can be real, and the proposed remedy can still advantage its authors. That is not a contradiction; it is a structural property of concentrated markets. Dismissing the capture argument requires engaging its structural basis, not its imputed motives.

Sacks also raises a motive argument — that product-liability exposure from a model implicated in a major cyber incident would be enormous, and that this exposure, rather than altruism, explains the timing. That is his argument about motive, not an established fact, and should be read as such. The liability dynamics he describes are real as a category; whether they explain the timing of this specific essay is his claim, not a finding.

The third prong of his critique is the verification mechanism — and this is the part that matters most structurally, because the evaluator commitment is the single concrete thing this week produced. Sacks raises questions about METR’s independence, pointing to what he characterises as ties to Anthropic’s investors and staff. That characterisation is his; it is not verified here, and METR has not publicly responded. But the structural point survives whether or not his specific characterisation holds. Independence is not a state that can be asserted; it is a property that must be engineered and demonstrated. In a field this small, professional ties of some kind between evaluators and the evaluated are close to inevitable. The answer is not to dismiss the question but to answer it with architecture: publication rights that survive the subject’s objection, funding that does not originate with the audited party, tenure protections, and recusal rules with teeth. None of that has been specified yet for either firm’s evaluator commitment, because neither firm has yet named an evaluator.

Where Each Tier Stands