Four CEOs Agreed in Public to Slow Down AI. Four Subscribers Say That Is Exactly the Problem.

Six days after Dario Amodei asked his rivals to slow down and three of them said yes in public, four people who pay for their chatbots asked a federal judge to call it a cartel.

The complaint, Buist v. Anthropic PBC, was filed Friday, Sept. 18, in the U.S. District Court for the Northern District of California and docketed as No. 3:26-cv-10693. It names Anthropic PBC, OpenAI OPCO LLC, SpaceXAI and Google LLC as defendants and alleges a single count: a horizontal agreement among competitors to restrain the rate at which their products improve, in violation of Section 1 of the Sherman Act. The plaintiffs want class certification for a proposed nationwide class of paid subscribers, treble damages, a declaratory judgment and an injunction. They have demanded a jury.

The named plaintiffs are Florida attorneys Charles Buist and Nick Spetsas, California attorney Cheyenne Hunt and California resident Christine Bullock. Three of the four hold paid subscriptions to all of Claude, ChatGPT, Grok and Gemini; Bullock pays only for Claude. They are represented by Trial Lawyers for Justice, whose lead attorney, Nick Rowley, is better known for catastrophic-injury verdicts than antitrust work. That shows in his framing. "AI will quickly spin out of human control and could kill us all if we allow AI safety and protocol ... to be controlled by private self-serving agreements between the world's most powerful 'for profit' technology companies," Rowley said in a statement carried by the Associated Press.

The Case Is Built From Press Releases

The 29-page complaint contains no leaked emails, no confidential witnesses and no meeting minutes. Its evidence is the public record of the week of Sept. 12. That Saturday, Amodei published "We Must Pace the Frontier," proposing limits on "the rate of unchecked AI progress," a three-stage plan whose second stage requires "industry-wide coordination," and a request that Washington "issue a narrow waiver for certain kinds of safety conversations." The same day, Elon Musk endorsed it in a four-word post, Sam Altman signaled agreement, and Demis Hassabis called it "the right path forward." The complaint's second paragraph says the agreement was "proposed in public, accepted in public, and confirmed in public."

The plaintiffs then reach back to July, when senior staff across the labs signed a statement acknowledging "intense competitive pressure not to unilaterally slow" development, and forward to Sept. 14, when Altman wrote that OpenAI did "not believe we need to wait for an antitrust exemption or legislation to begin the work." A day later OpenAI policy chief Chris Lehane confirmed the company had been working with Anthropic and Google DeepMind on safety for several weeks. The complaint treats that sequence as consciousness of antitrust risk: asking whether coordination was lawful, then proceeding without an answer.

The theory of harm is an output restriction dressed in software. "Plaintiffs challenge only what the antitrust laws forbid: an agreement among competitors about how fast their competing products will improve," the complaint states. "Congress has granted no exemption for that agreement." Subscribers, it says, keep paying the same monthly price for products that will improve more slowly than competition would deliver. It pleads, on information and belief, that the four defendants hold at least 80 percent of U.S. paid AI subscriptions. It does not identify a single delayed release, cancelled training run or downgraded model, and it concedes the full effect "has not yet manifested."

The plaintiffs are careful about what they are not challenging. Unilateral slowing, independent evaluators, safety research, compliance with government mandates and lobbying Congress for an exemption are all disclaimed. What they object to is the "shortcut" of agreeing to "substitute collective restraint for individual accountability." In the complaint's words: "The antitrust laws do not permit competitors to decide among themselves that competition is too dangerous."

Representatives for Anthropic, OpenAI, Google and SpaceXAI did not respond to requests for comment over the weekend.

Why It Matters

The doctrinal question is whether public agreement counts as agreement. Under Twombly, parallel conduct alone does not survive a motion to dismiss; four executives independently reaching the same conclusion on the same day is lawful conscious parallelism. The plaintiffs' answer is that this was not parallel; it was an offer and three acceptances, plus a working group Lehane confirmed on the record. Pascal Hetzscholdt, a competition and rights analyst who reviewed the filing, called that "meaningful factual enhancement" but added that the evidence is "not yet proof of the prohibited agreement." An AI safety working group, he noted, "is not intrinsically an output cartel."

If the case reaches the merits, the defendants' obvious move is to argue safety is procompetitive. That runs into National Society of Professional Engineers v. United States, where the Supreme Court refused to let a trade group suppress competitive bidding on the theory that competition produced unsafe engineering. The plaintiffs do not need to prove AI risk is fake; they need only show that private rivals, rather than Congress, decided how much competition was safe. Against that stands Allied Tube, which protects genuine, objective standard-setting, and a Justice Department official who said three days before the filing that cooperation on cybersecurity "does not appear anticompetitive."

The political context is not helping the defendants. FTC Chairman Andrew Ferguson said Sept. 15 that when companies come to Washington asking for regulation and an antitrust exemption at once, "all of my alarm bells go off," adding that "everyone should be deeply suspicious." Senator Josh Hawley stripped an AI antitrust carve-out from the defense authorization bill the same week and declared there was "no world" in which he would grant one. President Trump has called the slowdown a "conspiracy" and its safety rationale a hoax. Amodei's essay named the waiver as a precondition; every branch of government has now said no to it, and the complaint pleads that absence as an element of the offense.

The weak spot is damages. Filed six days after the alleged agreement, the suit cannot yet show a but-for world in which Claude or Gemini improved faster. Injunctive relief is the more realistic prize, and OpenAI's consumer terms include an arbitration clause and class waiver that may knock at least one defendant out before any judge reaches the Sherman Act.

What to Watch

Watch whether the three-lab working group keeps meeting; the complaint alleges it was active in mid-September, and continued sessions would be pleaded as continued conduct. Watch the first defense filings for arbitration motions and a Twombly challenge, likely within 60 days of service. Watch whether Google's newly launched DeepMind Institute, which has framed coordinated pacing as a research question, produces anything the plaintiffs can quote. And watch Brussels: Article 101 of the EU treaty expressly bars agreements that "limit or control technical development," and no European authority has yet said a word.