On September 18, four consumers filed a proposed class action in the US District Court for the Northern District of California against Anthropic, OpenAI, SpaceXAI and Google. According to MediaNama's report on the complaint, they allege the companies agreed to slow the development of competing AI products. They seek treble damages, an injunction and certification of a nationwide class of US consumers who bought paid subscriptions to ChatGPT, Claude, Grok or Gemini. Everything below concerns allegations. None has been tested in court, and the defendants' responses were not part of the reporting reviewed here.
The strongest case for the plaintiffs
Antitrust law does not carve out good intentions. Section 1 of the Sherman Act bars "every contract, combination, or conspiracy in restraint of trade," and the FTC's guide to the antitrust laws notes that the Clayton Act lets private parties harmed by violations seek triple damages and court orders. Courts have long treated agreements among competitors to restrict output as suspect, because rivals agreeing to make less of something is the classic way consumers get worse deals.
The complaint's factual theory is not frivolous on its face. It alleges that Anthropic, OpenAI and Google representatives formed a working group in July to develop an industry standards body. It also alleges that senior executives had acknowledged no company could slow down alone without losing customers and technological ground. Then, on September 12, Anthropic CEO Dario Amodei proposed "industry-wide coordination," including limits on training compute and on using AI to improve AI. The complaint notes that Elon Musk, Sam Altman and Demis Hassabis publicly endorsed the idea the same day. TechSpot's account adds that Amodei said the government would need to grant a narrow waiver for certain safety discussions, while Altman backed federal requirements but said work could proceed without waiting for legislation or exemptions.
If a court accepted the plaintiffs' framing, the fact that the participants believed they were acting for safety would not by itself be a defense.
Why the claim is still hard to win
A public proposal is not an agreement. Section 1 requires a contract, combination or conspiracy, and CEOs publicly endorsing a policy idea on the same day looks more like advocacy than a secret pact. The reporting describes a working group and a standards-body effort, not a signed commitment to cut compute or delay releases. Nothing in the reporting reviewed shows that any defendant actually slowed a product. Meanwhile the defendants continue to compete on models and subscription pricing.
As the FTC guide notes, the Supreme Court has held that Section 1 reaches only unreasonable restraints. Standards bodies, safety testing consortia and information-sharing on security threats are routine in other industries and are generally judged under the rule of reason, which asks whether the net effect harms competition. A plaintiff would also have to define a market, show harm to consumers in it, and show the alleged agreement caused that harm. That is a heavy lift when frontier model capabilities and subscription tiers have been expanding rapidly.
The real policy problem: nobody knows where the line is
The suit lands in a gap the agencies themselves created. In December 2024 the FTC and DOJ withdrew their Antitrust Guidelines for Collaborations Among Competitors, issued in April 2000, saying they "no longer provide reliable guidance." The FTC vote was 3-2, with Commissioners Andrew Ferguson and Melissa Holyoak dissenting. The agencies promised "vigorous antitrust enforcement on a case-by-case basis." In February 2026 the agencies opened a public inquiry into replacement guidance, with comments due April 24, 2026. FTC Chairman Ferguson said that businesses need "transparency and predictability" and that the withdrawal had left "millions of businesses in the dark." No replacement guidance appears in the reporting reviewed here.
That matters because safety collaboration among frontier developers is precisely the kind of conduct where the boundary between legitimate standard-setting and output restriction is unclear. Private treble-damages litigation is a poor tool for drawing that boundary. Its outcomes are set by a handful of district judges, discovery can be used as leverage, and the threat of trebled damages across a nationwide class can chill lawful cooperation long before any court reaches the merits.
What proportionate policy looks like
The evidence-based response is not to dismiss the suit's concern or to grant the industry a blanket exemption. It is to make the rules legible:
- Finish the guidance. The agencies should publish replacement collaboration guidance that distinguishes safety testing, common evaluation protocols and incident reporting from agreements on output, pricing, release timing or compute budgets.
- Use existing tools for clearance. Companies uncertain about a specific arrangement can seek agency review before acting, rather than discovering the answer in a class action.
- Keep transparency high. Standards work conducted openly, with outside evaluators and published outputs, is easier to defend and harder to portray as a cartel.
- Be skeptical of limits on compute as a safety tool. Whether or not it is legal, an industry-wide cap agreed among the largest incumbents would also raise entry barriers for smaller labs and open-source developers. That is a competition cost that the proposal's supporters should confront directly.
The last point cuts in an uncomfortable direction for the industry. Even if the complaint fails on the law, a mechanism that lets the current market leaders jointly decide how fast the frontier moves deserves scrutiny from people who care about innovation. Regulators and courts can hold both views: that safety cooperation should be legally possible, and that it should not become a way for incumbents to fix the pace of competition.
What to watch
The first tests will be a motion to dismiss on whether the complaint pleads an actual agreement rather than parallel public statements, and any response from the FTC or DOJ. A dismissal would leave the guidance gap open. A denial would push discovery into the July working group's communications, which would tell the public far more about what was and was not agreed than the complaint's summary does.