Class action alleges OpenAI, Anthropic, Google and SpaceXAI coordinated to slow AI releases

A cartel case built from tweets and essays

Four paying subscribers of ChatGPT, Claude, Grok and Gemini filed a class action in the US District Court for the Northern District of California on September 18, 2026, accusing Anthropic, OpenAI, SpaceXAI and Google of agreeing to slow the pace at which their consumer AI products improve. According to a Netics Labs analysis published on dev.to, the suit — Buist et al. v. Anthropic PBC et al., No. 3:26-cv-10693 — frames the alleged coordination as a horizontal restraint under Section 1 of the Sherman Act. The theory of harm is that an agreement to slow improvement reduces the quality subscribers receive for the price they pay, which the complaint characterizes as an overcharge.

What makes the filing unusual, Netics argues, is not the legal theory — agreements among competitors to restrict output are a familiar antitrust category — but the evidence. The complaint is assembled almost entirely from material the defendants published themselves.

One day of public messaging

The sequence at the center of the filing took place on September 12. That morning, Anthropic CEO Dario Amodei published an essay titled We Must Pace the Frontier. Within roughly an hour, Elon Musk quote-posted it with "Dario is right." Later the same day, Sam Altman wrote that he agreed with Amodei about pacing the frontier, and the complaint says Demis Hassabis endorsed the direction while leaving implementation details open. Three days later, per the Netics write-up, OpenAI's policy chief confirmed the talks had been running for weeks.

The complaint's summary of this record, as quoted by Netics: the agreement was "proposed in public, accepted in public, and confirmed in public."

What the labs allegedly agreed to

Paragraph by paragraph, Netics reports, the filing pleads the elements of a deal: the parties (the frontier labs), the term (a slower rate of capability advancement than each would choose alone), and the mechanisms — shared limits on training compute, on training runs and on the internal use of AI to improve AI, plus capability checkpoints. Defection would be policed, the filing alleges, through "verifiable pacing through embedded evaluators."

The complaint also cites Amodei's own acknowledgment that the government "would need to issue a narrow waiver for certain kinds of safety conversations," and answers it flatly: no such waiver exists, Congress has enacted no exemption, and no agency compelled the conduct.

On market power, the filing puts the four defendants at "at least 80%" of paid US consumer subscriptions to general-purpose frontier generative-AI assistants — a figure pleaded on information and belief and framed as a floor. Netics highlights the strategy here: by characterizing the restraint as naked, meaning a horizontal agreement about output with no plausible efficiency justification, the plaintiffs can lean on long-standing doctrine that condemns such agreements without a full market analysis. Whether a court treats cross-lab safety coordination as a naked restraint or as something closer to standard-setting with competitive benefits is, in Netics' reading, the entire case in one sentence.

The gap in the filing

The complaint also contains its weakest point. Across its 29 pages, per Netics, it identifies no delayed release, no cancelled model and no altered training run — the filing itself concedes the effect on shipped products has not yet appeared. Under the rule of reason that gap would likely be fatal; under per se treatment it is merely awkward. The four public acceptances also differ in weight: Musk's was four words, Google's endorsed a direction with details left open, and Amodei's essay explicitly invited government mediation.

Why it matters

The case is, in Netics' framing, a test of how much legal exposure public safety coordination can create — for AI labs and for any industry that coordinates safety in the open. Every sentence of the evidence was public advocacy when it was made; the same record that signaled responsibility to buyers and regulators is now the backbone of a federal antitrust claim, and Amodei flagged exactly that risk in his essay.

For buyers, Netics flags three things to watch. First, the operational terms of the embedded-evaluator program, which discovery would scrutinize first and which customers would want to see anyway. Second, the complaint entrenches the idea that paid AI subscriptions are bought for the pace of improvement, turning release cadence into a quality expectation. Third, settlement risk: a four-defendant class action with treble damages and a jury demand rarely ends at the motion-to-dismiss stage even on thin merits. The honest reading, Netics concludes, is that the case could go either way — naked-restraint treatment means an expensive discovery fight over the labs' own words, while a finding of coordinated standard-setting narrows the case to whether the coordination was reasonable, a fight the labs may well win.