Four Labs Agreed to Slow AI Down. Now They're Being Sued for Agreeing.

We build agentic products on top of models from most of the labs named in this lawsuit, so the legal theory here isn’t abstract to us: it argues that safety-motivated coordination between the companies whose APIs we build on top of is itself illegal. On September 19, four paying subscribers to ChatGPT, Claude, Grok, and Gemini filed a proposed class action in the U.S. District Court for the Northern District of California against Anthropic, OpenAI, xAI, and Google, arguing the four labs illegally agreed to slow down AI development and that the agreement made their subscriptions worth less than advertised, according to Fortune and CNN Business.

What supposedly happened on September 12

The alleged coordination traces to one week of public statements, not a signed agreement:

• Anthropic CEO Dario Amodei published an essay calling for industry-wide deceleration: “We must slow the pace at which we improve the capabilities of AI models.” His concrete asks, since nicknamed “Pace the Frontier” in industry coverage, included Anthropic unilaterally giving external evaluators employee-level access to its safety procedures and incident reports, urging other labs to do the same, and calling on democratic governments to set binding safety standards — chip export restrictions to adversarial nations, mandatory testing requirements for frontier models — per Axios.
• Sam Altman publicly agreed with the need to slow frontier development and committed OpenAI to giving external evaluators similar access.
• Elon Musk endorsed the position directly, posting “Dario is right” — notable because, as the lawsuit points out, Anthropic is also a major customer of Musk’s data-center capacity.
• Demis Hassabis is named as a defendant via Google; the specific nature of his public response wasn’t detailed in the coverage we could find.

Why the legal theory is unusual

Every prior antitrust fight involving AI labs has argued the opposite of this one — that they’re moving too aggressively, foreclosing markets, or making killer acquisitions. This lawsuit flips it: it argues that publicly, simultaneously agreed-upon restraint is itself the anticompetitive act. The theory didn’t come from nowhere — legal commentary anticipating exactly this argument ran under the headline “Move Slow and Collude: The Antitrust Problem With Pacing AI” two days after Amodei’s essay and five days before the suit was filed. Its core point: an express, public, cross-company agreement that competitors’ technological progress “should be slower than competition would otherwise produce” looks structurally like a textbook horizontal restraint — the kind of coordinated output restriction antitrust law treats skeptically even when the stated motive is safety.

The theory has an obvious hole, though, and it’s the one that will decide whether the case survives a motion to dismiss: antitrust law requires proof of an actual agreement, not just parallel public statements reacting to the same essay. Some coverage of the same week framed it more loosely — “Three AI Rivals Agree: Slow the Frontier Down” — as a PR alignment rather than a negotiated pact. Altman agreeing with Amodei’s essay and Musk posting three words of support is a long way from a documented pact to restrict output. Whether that gap matters legally, or whether the public, coordinated timing of the statements is itself enough to infer agreement, is exactly what a judge will have to decide.

The tension we’re not going to paper over

If this theory gets any traction at all, it creates a genuinely awkward bind. Policymakers, AI-safety researchers, and the Global Call for AI Red Lines campaign — Nobel laureates included — have spent the past year pushing labs toward exactly this kind of coordinated, public commitment to caution. A ruling that treats that coordination as a horizontal restraint would tell every lab that the safe legal move is to make safety commitments privately, unilaterally, and without acknowledging what a rival is doing — the opposite of the transparency the push for pacing was trying to build. On the other side, “we did this for safety” has not historically been a winning defense against a garden-variety antitrust claim of express competitor coordination, and courts have generally not been sympathetic to public-interest justifications for it in other industries. Both of those things can be true at once, which is what makes this an actual open legal question rather than a story with an obvious right answer.

What we’re watching

None of the four defendants had responded to requests for comment as of the most recent coverage we found. The suit doesn’t yet specify requested remedies beyond representing a proposed nationwide class of paid subscribers. From here, the things that will actually tell us how serious this is: whether the case survives an early motion to dismiss on the agreement-versus-parallel-conduct question, and whether labs that didn’t join the September 12 statements — Meta, Mistral, DeepSeek — get invoked by either side as a comparison class showing the market could have moved faster absent the alleged agreement. If it does survive, every future joint safety statement from a frontier lab is going to get run past a lawyer before it gets run past a comms team — including, presumably, at the labs we build on top of.