Four consumers filed a class action in the Northern District of California on 18 September against Anthropic, OpenAI, SpaceXAI and Google. Buist v. Anthropic PBC alleges a horizontal agreement to slow the rate at which the four improve competing products, contrary to section 1 of the Sherman Act. The case matters because the evidence rests on the companies' own public statements rather than internal documents.
What the complaint alleges
The filing, reported by Bloomberg Law, seeks treble damages and an injunction, and demands a jury. The pleaded term is a slower rate of capability advancement than each company would choose alone. The mechanisms alleged are limits on training compute and training runs, restrictions on using AI to improve AI, and capability checkpoints. The policing mechanism alleged is the embedded evaluator proposal itself, described as verifiable pacing — the same programme Anthropic has begun implementing with Accenture. The claimed 80% share of US paid subscriptions is pleaded on information and belief and as a floor rather than a measurement.
The complaint's factual sources are named executives' public statements and press reports. Amodei's essay said government mediation would help and that a narrow waiver would be needed for certain safety conversations; the complaint treats that essay as the offer. Two days later Altman said OpenAI did not intend to wait for an antitrust exemption or legislation, which the filing treats as acceptance. The following day OpenAI's policy chief Chris Lehane confirmed the company had already been working with Anthropic and Google DeepMind for several weeks. Paragraph two of the complaint states the agreement was proposed in public, accepted in public, and confirmed in public.
There are no internal documents, no confidential witnesses and no meeting minutes in the filing. The four acceptances are not equivalent: Musk's was four words on social media, and Google's was an endorsement of the direction with implementation details left open. The complaint files the sequence under consciousness of antitrust risk — asking whether the conduct was lawful is pleaded as proof the companies understood it as coordination, and proceeding anyway as proof of intent. The plaintiffs also plead the restraint as naked, which would make market definition unnecessary, and offer the market only in the alternative.
What this means for business
The legal move is narrower than it first appears. The filing argues that competition being dangerous is not a defence a court may weigh, because that judgment belongs to Congress. It does not allege the safety concerns are invented; its pretext argument is structural, holding that every stated safety objective can be pursued unilaterally, through independent evaluators, or through regulation. The plaintiffs say they take AI risk seriously, and paragraph eleven describes independent safety investment and welcoming government oversight as the right path. They disclaim any challenge to unilateral slowing, independent evaluators, lawful safety research, compliance with government requirements, or petitioning Congress, including petitioning for an exemption.
For companies buying or building on frontier models, the practical question is what a court could order if the plaintiffs prevail. Treble damages and an injunction are the stated remedies, and an injunction could touch the pacing programmes rather than the models themselves. The complaint concedes that the full effect on released products has not yet manifested: no slowed release, cancelled model or altered training run is identified anywhere in twenty-nine pages. That gap matters for procurement teams, because it means no current contract or roadmap has yet been shown to be affected by the alleged conduct.
The antitrust exposure was discussed openly before anyone filed. A statement from employees across the labs in July said each faced intense competitive pressure not to unilaterally slow, which the complaint now uses as evidence of both motive and action against self-interest. Article 101 of the EU treaty prohibits agreements between undertakings that restrict competition and contains no safety exemption a company can grant itself, so a coordinated pacing arrangement among frontier developers would face the same analysis in Brussels. There is no European case, and none has been signalled. Beijing called the slowdown plan a Cold War script, which leaves coordination contested at both ends.
Watch whether the working group meets again. The complaint alleges it was still meeting in mid-September, and continuing now would be read as continuing the conduct. Watch the answers as well: these are allegations filed the day before this was written, no defendant has responded, and the companies have strong arguments that public advocacy and unilateral safety work are lawful. For business, the marker is whether any vendor changes a published release schedule or training plan in response to the case.
