Confirmed in a primary source 18 Sept 2026 Law

Four people who pay for ChatGPT, Claude, Grok and Gemini sued all four companies for agreeing to slow down.

On September 18 four subscribers filed a class action in San Francisco federal court against Anthropic, OpenAI, SpaceXAI and Google, arguing that the companies' public agreement to slow the pace of AI improvement is an illegal agreement among competitors to make their products worse, and asking for three times the money the class lost plus a court order banning the arrangement. The caveat that matters is that nothing in a complaint has been tested: no defendant has answered, no judge has ruled, and the defence that safety is not a product feature has not been heard yet.

Until now

The pacing agreement was a policy argument, to be settled by regulators, by Brussels, or by the antitrust exemption the companies said they would ask Washington for.

Now

Treat the pacing agreement as a named defendant in a federal court that can reach it before any regulator does, and read every public safety commitment as a document a plaintiff can quote back.

What changed is the forum. Edition 020 carded the week the labs agreed in public to pace themselves and promised outside evaluators desks and badges, and the reading then was that the next move would be the companies going to Washington to ask permission to coordinate. Altman said so himself on September 14: OpenAI would not wait for an antitrust exemption or legislation before starting the work with colleagues across the industry. Four subscribers took him at his word and got to court first. The complaint needs no discovery to state its case, and that is its design. Every act it alleges happened on a public timeline, in essays and posts the companies published themselves, which is why its second paragraph says the agreement was proposed in public, accepted in public and confirmed in public.

The legal question underneath is narrow and old. Section 1 condemns agreements among competitors to restrict output, and the complaint argues that agreeing to improve your products more slowly restricts output in the one dimension where these four firms actually compete, which is how fast the models get better. The answer will be that safety is not a product feature and that coordination about catastrophic risk is not the restraint the antitrust laws were written to stop, and it will have to be made to a jury the plaintiffs have already demanded. The plaintiffs have pre-empted the obvious reply: they say plainly they are not against regulation or against an antitrust exemption, only against four companies setting the pace among themselves with neither one in place. Treat the pacing agreement as a named defendant in a federal court that can reach it before any regulator does, and read every public safety commitment as a document a plaintiff can quote back.

How it got here

  1. Jul 2026

    Representatives of OpenAI, Anthropic and Google DeepMind begin meeting in a working group, according to the complaint, which cites OpenAI's global policy chief.

  2. Sep 12

    Amodei publishes the pacing essay. Musk, Altman and Hassabis agree the same day. Anthropic and OpenAI both promise embedded evaluators. Carded in Edition 020.

  3. Sep 14

    Altman says AI progress should be slower than it otherwise could be, and that OpenAI will not wait for an antitrust exemption or legislation. The class period is already running.

  4. Sep 18

    The complaint is filed in San Francisco. One count, four defendants, a nationwide class and a jury demand.

  5. Who has to react

    The general counsels of all four defendants, who have twenty-one days from service to answer or move; the Antitrust Division and the Federal Trade Commission, which now have a private complaint laying out the theory before either has said whether coordination on safety is lawful; every other plaintiffs' firm in the country, because the first complaint in an antitrust wave is a template; Anthropic's bankers, with a November listing and a securities disclosure that now has to describe this; the embedded evaluators themselves, named in paragraph 66 as the enforcement arm of the alleged conspiracy rather than as a safety measure; the European Commission, which endorsed pacing language in its State of the Union address on September 16; and every lab lawyer who has been asked to review a safety blog post this week.

The evidence

ObservedFriday September 18, 2026. Charles Buist and Nick Spetsas of Florida and Cheyenne Hunt and Christine Bullock of California filed a 29-page class action complaint in the United States District Court for the Northern District of California, San Francisco Division.

The whole observation, as recorded

Case number 3:26-cv-10693, assigned to Magistrate Judge Nathanael M. Cousins, cause of action 15:1 antitrust litigation, nature of suit 410, filing fee 405 dollars, receipt ACANDC-22486761, jury demanded. Defendants: Anthropic PBC, OpenAI OpCo LLC, SpaceXAI LLC and Google LLC. Lead counsel is Nicholas C. Rowley of Trial Lawyers for Justice, with Andrew T. Tutt and R. Stanton Jones on the signature block. One count: Section 1 of the Sherman Act, 15 U.S.C. 1. Relief sought: damages trebled under Section 4 of the Clayton Act, 15 U.S.C. 15; an injunction under Section 16, 15 U.S.C. 26, against any horizontal agreement covering development rates, compute limits, coordinated delays or capability checkpoints; fees, interest and a jury. The proposed nationwide class is everyone in the United States who bought a paid individual consumer subscription to ChatGPT, Claude, Grok or Gemini directly from a defendant from September 12, 2026 onward, with four company-specific subclasses. The evidence pleaded is entirely public: Dario Amodei's September 12 essay We Must Pace the Frontier and its line that we must slow the pace at which we improve the capabilities of AI models; Elon Musk's reply about an hour later, Dario is right; Sam Altman's I agree with Dario that we need to pace the frontier; Demis Hassabis calling the essay the right path forward and tying it to the industry-wide standards body he had proposed two months earlier; Altman two days later saying progress should be slower than it otherwise could be and that OpenAI would not wait for an antitrust exemption or legislation; and OpenAI's global policy chief confirming a working group of company representatives meeting regularly since July. Paragraph 66 of the complaint names embedded evaluators as the enforcement mechanism, quoting Amodei that they would make verifiable pacing practical.

Established

  • A complaint alleging that Anthropic, OpenAI, SpaceXAI and Google agreed to slow the improvement of their products, in violation of Section 1 of the Sherman Act, was filed in the Northern District of California on September 18, 2026, as case 3:26-cv-10693.
  • The plaintiffs are four named individuals who subscribe to the paid tiers, and they seek treble damages, an injunction against horizontal agreements on development rates, compute limits, coordinated delays and capability checkpoints, and a jury.
  • The complaint pleads only public statements: the September 12 essay and replies, Altman's September 14 remarks, and a July working group described by OpenAI's own policy chief.
  • The complaint names embedded evaluators as the mechanism that would verify compliance with the alleged agreement.

Not established

  • That an agreement exists in law. Public agreement with a public proposal is evidence a plaintiff can plead, and whether it amounts to a contract, combination or conspiracy under Section 1 is exactly the contested question.
  • That the case survives. No defendant has answered, and motions to dismiss in Section 1 cases turn on pleading standards that have sunk better-supported complaints.
  • That subscribers were harmed. The theory is that a slower product is a worse product and therefore an overcharge, and it has not been tested by any economic evidence on this docket.
  • That the government agrees. No federal or state enforcer has opened a public investigation into the pacing agreement, and none is named in the complaint.

Why it weighs this much

Signal strength
CostModerate
CommitmentHigh
ReversibilityHard
PersistenceYears

Cost is moderate rather than high: the filing fee is 405 dollars and the real spending is a contingency firm's time, but that firm has now taken on four of the best-funded defendants in the country with no public co-counsel and no government case to ride behind. Commitment is high because a filed federal complaint with a jury demand is not a press release. It binds the plaintiffs to litigate, it exposes counsel to sanctions if the allegations are frivolous, and it puts a public document on a docket that every other firm can copy. Reversibility is hard for the same reason: a voluntary dismissal would not remove the complaint from the record or stop the next plaintiff. Persistence is years because antitrust class actions of this shape run three to six years before any merits ruling, and the class period is open-ended, running until the effects of the alleged conduct cease.

What to watch

  • weakens The Section 1 claim is dismissed on the pleadings without leave to amend, or the complaint is voluntarily dismissed, or the Justice Department or Federal Trade Commission publishes guidance or a safe harbour permitting coordination on AI safety, by June 30, 2027.
  • strengthens A court denies a motion to dismiss the Section 1 claim, or a state attorney general, the Federal Trade Commission or the Justice Department opens a public investigation or files its own case over the pacing agreement, by March 31, 2027.

The words

Section 1 of the Sherman Act
The 1890 law making an agreement between competitors to restrain trade illegal. It covers agreements about how much is produced or how good it is, not only agreements about price.
Horizontal agreement
An agreement between direct rivals, as opposed to one between a company and its supplier. Courts treat horizontal agreements far more harshly, and some are illegal without any inquiry into their effects.
Treble damages
Under the Clayton Act a winning antitrust plaintiff collects three times what it actually lost. This is what makes private antitrust cases worth a law firm's time.
Class action
One suit brought on behalf of a defined group. Here the group is everyone in the United States who has paid for one of the four subscriptions since September 12.

Your call

Which explanation do you believe?

A call is final. Other people's calls show after yours.

8 contributions. Thank you.

You have built
a record here.

Keep it: create an account to verify your contributions and put your name to the answers you choose. Your record stays yours.

Keep my record ↗
Optional: help show whose perspectives are missing
Kept with your answers under a random id, never with your name. Used for one thing: showing you, on your record, where you stand next to people like you, once at least 20 of them have answered.

Add your signature

Stand behind your judgment.

Create an account to attach your name to this answer. Your other answers show no name unless you sign them too.

Sign up to add your name ↗

Already have an account?