Four Subscribers Sue OpenAI, Anthropic, Google and SpaceXAI Over an AI Slowdown Pact

TL;DR

On Friday, September 18, four people who pay for ChatGPT, Claude, Grok and Gemini filed Buist et al. v. Anthropic PBC et al., No. 3:26-cv-10693, in the U.S. District Court for the Northern District of California. The claim is Section 1 of the Sherman Act. The theory is that Anthropic, OpenAI, SpaceXAI and Google DeepMind agreed among themselves to slow the rate at which their competing products get better, which is an output restriction wearing a lab coat. The unusual part is the evidence. There are no leaked Slack logs and no whistleblower. The complaint is assembled almost entirely from things the defendants published on purpose, some of it only six days before it was filed.

What actually got filed

The named plaintiffs are Charles Buist and Nick Spetsas of Florida and Cheyenne Hunt and Christine Bullock of California, each a paying subscriber to at least one of the four assistants, suing on behalf of a proposed nationwide class. Nicholas C. Rowley of Trial Lawyers for Justice leads the plaintiffs' team. They want treble damages, an injunction barring horizontal agreements on development pace, compute limits and release coordination, and a jury.

The relevant market they plead is narrow and deliberate: US paid consumer subscriptions to general-purpose frontier AI assistants, meaning the paid tiers of ChatGPT, Claude, Grok and Gemini. On information and belief, the complaint puts the four defendants at no less than 80 percent of that market.

The theory: pacing is an output restriction

Antitrust law has never struggled with competitors who agree to make fewer widgets. That is the oldest cartel in the book, and courts treat it as illegal per se, meaning no one gets to argue it was actually good for everybody. The complaint's move is to claim that a rate of improvement is also an output. A bakery cartel agreeing to bake fewer loaves and a lab cartel agreeing to make next year's model only a bit better than this year's are, on this reading, the same arrangement in different clothes.

The doctrinal support for that move sits in cases where courts refused to let defendants argue that competition itself was the problem, including National Society of Professional Engineers v. United States, where the Supreme Court threw out an engineering society's public-safety justification for banning competitive bidding, and NCAA v. Board of Regents. The complaint's one-line version of the argument: "The antitrust laws do not permit competitors to decide among themselves that competition is too dangerous."

The injury theory is mundane by comparison, which is probably the point. Subscriptions are sold on access to each company's best models and on the promise that those models keep improving. Slow the improving and subscribers pay the same price for less. That is an overcharge, and overcharges are what antitrust damages are built on.

The evidence is a blog post

Most Section 1 cases live or die on whether plaintiffs can plead enough to get into discovery and go looking for the agreement. This one arrives with the agreement already printed.

The anchor is Anthropic CEO Dario Amodei's September 12 essay We Must Pace the Frontier, whose central ask is one sentence long: "We must slow the pace at which we improve the capabilities of AI models." Within two and a half hours, Elon Musk, Sam Altman and Demis Hassabis had each publicly agreed. The essay also proposes the mechanisms the plaintiffs now want enjoined: capability checkpoints, and pacing "based on limiting the ingredients that go into frontier models, such as training compute, the nature of training runs, or internal use of AI to improve AI."

The detail that turns a policy debate into a pleading is the waiver. Amodei wrote that the US government does "need to issue a narrow waiver for certain kinds of safety conversations." That is a lab CEO saying, in public, that this category of conversation has an antitrust problem unless someone clears it. Three days later, on September 15, OpenAI global policy chief Chris Lehane told reporters in Washington that the talks with Anthropic and Google DeepMind had been running for several weeks and that no waiver was needed. Both statements are now in the complaint, doing opposite jobs for the same plaintiffs: one establishes that the participants knew the line existed, the other establishes that they walked over it anyway.

Whatever else you think of the case, the tradecraft is remarkable. If you are going to run a cartel, the traditional venue is a hotel suite in Geneva, not a personal blog with an RSS feed.

What this does to safety coordination

The defendants have not answered, and the obvious first move is a motion to dismiss arguing that parallel public statements of intent are not an agreement. Four CEOs saying the same thing in one afternoon is a fact pattern courts see often, and "we all independently reached the same conclusion" is a real defense. Lehane's confirmation that actual talks were under way is what makes it harder than usual.

For anyone building on these models, the practical stakes are not the damages. They are the injunction. The relief requested would bar the four from agreeing on development pace, on compute limits, and on release timing, which is precisely the machinery Amodei's plan runs on. Steps two and three of that plan, cross-lab standards and capability checkpoints, do not survive an order like that unless Congress creates a safe harbor first.

There is a version of this where the case is good for safety and a version where it is corrosive, and they are hard to tell apart from here. The corrosive version: labs keep coordinating and stop saying so, which makes the coordination less observable without making it less real. The useful version: the legitimate route becomes statutory. The FRONTIER Act (H.R. 9925), from Reps. Lori Trahan and Jay Obernolte, licenses independent verification organizations to sit inside frontier labs and assess them, a government-supervised structure rather than a private handshake. Lehane has said OpenAI backs that provision. A statute that tells labs what they must do is not an agreement among labs, which is the whole difference antitrust cares about.

The caveats, straight

• Everything above is an allegation. Nothing has been proven, none of the four had responded publicly as of the weekend, and the case is at the earliest possible stage.
• The 80 percent market share is pleaded on information and belief, not established. Market definition is where cases like this usually get decided.
• Per se treatment is the plaintiffs' preferred framing, not a given. If a court applies the rule of reason instead, the defendants get to argue the safety benefits, and the case gets a great deal longer.
• Quantifying what a class of subscribers lost because a model improved more slowly than it otherwise would have is not a normal damages exercise. It may be the hardest part of the whole thing.

Key Takeaways

• Buist et al. v. Anthropic PBC et al., No. 3:26-cv-10693, was filed September 18, 2026 in the Northern District of California by four paid subscribers against Anthropic, OpenAI, SpaceXAI and Google.
• The claim is Section 1 of the Sherman Act on an output-restriction theory: agreeing to slow the rate of product improvement is treated like agreeing to cap production.
• The complaint is built from public statements, including Amodei's September 12 essay and OpenAI policy chief Chris Lehane's September 15 confirmation that cross-lab talks had been running for weeks.
• Amodei wrote that such coordination needs a narrow government antitrust waiver. Lehane said it does not. Both lines now appear in the same complaint.
• Plaintiffs seek treble damages plus an injunction against agreements on development pace, compute limits and release coordination, which is the exact machinery of the pacing plan.
• The statutory alternative already exists in draft: the FRONTIER Act would license independent verifiers inside labs, replacing a private handshake with a government-supervised one.

Sources: The Next Web, The AI slowdown pact is now a Sherman Act lawsuit, Unite.AI, Consumers Sue Anthropic, OpenAI, SpaceXAI and Google Over Alleged AI Pact, CBS News, PBS NewsHour, Forkast, Bloomberg Law, Latin Times, Dario Amodei, We Must Pace the Frontier, TechCrunch, H.R. 9925, FRONTIER Act