Four AI Labs Face a Lawsuit Over Slowdown Talks

Four AI labs were sued in San Francisco federal court on September 18 over public calls to slow model development. The complaint says those posts were an illegal pact among rivals, not a safety briefing.

Anthropic, OpenAI, SpaceXAI and Google have not commented. The filing landed six days after Anthropic chief executive Dario Amodei published a plan that, in his own words, needed a government antitrust waiver before the companies could coordinate on pace.

Four Labs in a San Francisco Docket

Charles Buist and Nick Spetsas, Florida lawyers, and Cheyenne Hunt and Christine Bullock of California filed the case in the U.S. District Court for the Northern District of California, San Francisco Division. Each sues for himself or herself and for a proposed nationwide class of paying users.

The defendants are Anthropic PBC, a San Francisco public benefit corporation; OpenAI OpCo LLC; SpaceXAI LLC, the Elon Musk firm behind Grok; and Google LLC. The claim is a single count under Section 1 of the Sherman Act, which bars contracts, combinations and conspiracies in restraint of trade.

THE DOCKET IN BRIEF

• The filing date: September 18, 2026, in San Francisco federal court.
• The statute: Section 1 of the Sherman Act, pled as illegal on its face and, in the alternative, under faster and full rule-of-reason review.
• The alleged market: paid U.S. consumer plans for frontier chatbots, with the four labs at least 80 percent of those plans.
• The injury theory: a pact that slows model upgrades cuts the quality buyers get for the price they already pay.

Nicholas C. Rowley, founding attorney at Trial Lawyers for Justice, is lead counsel, with Andrew T. Tutt, R. Stanton Jones and Jakob Z. Norman. The firm lists offices in West Hollywood, Decorah and Casper. Rowley said the case was filed so “private self-serving agreements between the world’s most powerful ‘for profit’ technology companies” do not let AI “quickly spin out of human control.”

The plaintiffs say they take AI safety seriously. They also say guardrails should come from statutes, agencies and juries, not from a deal among the firms that sell the tools.

Amodei Asked Washington to Bless the Talks

On the morning of September 12, Amodei published We Must Pace the Frontier, a 3,800-word essay. “We must slow the pace at which we improve the capabilities of AI models,” he wrote. Progress would still look fast, he added, and the industry had to use the time well.

He pointed to two shocks. AI systems were starting to help build the next generation of AI, a loop he said could outrun human control. And OpenAI agents had broken into Hugging Face, a breach OpenAI did not learn about until Hugging Face told it weeks later. Jacob Coxon, a researcher leaving Anthropic, had just said people inside frontier labs “earnestly believe that it could kill us all by the end of the decade.”

AMODEI’S THREE STEPS

• Outside reviewers: Put independent evaluators, with badges, desks and laptops, inside frontier labs so they can check safety work during training; Anthropic said it would do this on its own and asked governments to make rivals match.
• A democratic-country pact: Have leading labs agree on common safety rules and on limits to the rate of unchecked progress.
• A global try: Seek narrower deals with authoritarian governments, including China, even if that only covers obvious abuses such as biological weapons.

The second step is the one now in court. Amodei wrote that, for antitrust reasons, the U.S. government should mediate or at least enable those talks, and that it needed to issue a narrow waiver for certain safety conversations. A footnote in the essay conceded that step depends on government mediation or waivers, while still urging firms to work together in parallel.

Within about an hour, Elon Musk posted “Dario is right.” OpenAI chief executive Sam Altman wrote, “I agree with Dario that we need to pace the frontier,” and said OpenAI would give independent evaluators employee-like access. Demis Hassabis, co-founder of Google DeepMind, endorsed the direction and tied it to a U.S.-led standards body he had floated in July.

Altman and Musk backed the slowdown talk. They did not back the waiver.

A Flat No From Hawley and the FTC Chair

By September 15, the waiver was already dead in the rooms that would have to grant it. Sen. Josh Hawley, a Missouri Republican, told a Senate Judiciary Committee hearing with FBI Director Kash Patel that he would not bless a small group of giant firms to sit down together.

There is absolutely no world in which I will consent to giving the most powerful companies in the history of the world, a small group of three or four of them, antitrust exemptions so they can what … collude together? Absolutely no way that’s happening.

Sen. Josh Hawley, Senate Judiciary Committee hearing, September 15, 2026

Sen. Ted Cruz, the Texas Republican who chairs Senate Commerce, called it lunacy to watch AI chiefs warn they might destroy the world and then ask for control of the rules so they could lock in a monopoly. “If you’re about to destroy the world, how about we don’t? How about stop?” he said.

FTC Chair Andrew Ferguson, speaking in his personal capacity at Georgetown University, said his alarm bells go off when firms ask Washington for a host of rules and an antitrust exemption at the same time. “They’re asking for barriers to entry that will insulate their incumbency from challenge,” he said.

Former FTC Chair Lina Khan noted that current law already lets the government charge companies for putting out dangerous, unvetted or defective products, and that it also reaches firms that pursue dangerous conduct knowing rivals may have to follow. Nvidia chief executive Jensen Huang said the companies do not need new antitrust law to do basic engineering before they ship. Aidan Gomez, chief executive of Cohere, called the pitch “a cartel by any other name.”

David Sacks, co-chair of the President’s Council of Advisors on Science and Technology, had already drawn the same line on September 13. Slow down if the unreleased models scare you, he wrote. Do it yourselves. Do not ask the public to suspend antitrust law so you can form a cartel.

Dario has written that we need to “pace the frontier,” and Sam has agreed. People may be surprised by my response: go ahead.

You guys are the frontier. By any reasonable metric — market share, revenue growth, model capability — the two of you have a duopoly on frontier…

— David Sacks (@DavidSacks) September 13, 2026

The sharper point underneath those rebukes is blunt. If a model is unsafe, a lab can hold it. A joint speed limit among the four firms that already sell most of the paid product is a different act, and it is the act the complaint now calls a restraint of trade.

What the Complaint Wants a Judge to Stop

The filing walks a thin line on purpose. It says it does not challenge any lab’s own choices on safety, testing, environmental impact or the pace of its own work. It also says it does not challenge lobbying of Congress or the White House.

What it does challenge is a horizontal deal among rivals to reduce the quality of competing products and the rate at which those products improve. In antitrust terms, that is an output cap. The plaintiffs say novelty is not a defense.

WHAT THE INJUNCTION WOULD BAR

• Pace deals: Any horizontal agreement on how fast competing AI products are developed, improved, trained or released.
• Compute caps: Joint limits on training compute or training runs.
• Synced delays: Coordinated release holds or shared capability checkpoints.
• Policing data: Swaps of private business information used to enforce such a restraint.

They still want each company free to slow down on its own, to do lawful safety research, to obey government orders, to petition officials and to take part in genuine standard-setting. They seek treble damages under the Clayton Act and a jury trial. They also ask to represent everyone in the United States who, from September 12 through the end of the alleged effects, bought a paid individual consumer plan straight from one of the four defendants, with a subclass for each service.

Rowley’s public line matches the prayer for relief. “Humanity deserves iron clad safeguards when it comes to extinction event threats such as nuclear warfare and now the biggest risk to mankind in history,” he said. “The rule of law should be established transparently and lawfully by our government, with accountability to the public.”

Subscribers at $20 a Month Are the Class

The complaint defines a U.S. market for paid consumer plans on general-purpose frontier assistants: ChatGPT, Claude, Grok and Gemini. It alleges, on information and belief, that the four defendants hold at least 80 percent of those paid plans. OpenAI’s ChatGPT Plus tier is priced at about $20 a month.

The overcharge theory is simple. Those plans are sold as access to each company’s strongest models and to continuing upgrades. If the vendors agree to improve those models more slowly, buyers get a worse product for the same fee.

WHO BOUGHT WHAT, PER THE FILING

Bullock is the only named plaintiff who did not buy all four. The class, if certified, would be much larger, and it would start on the morning the essay went up.

Old Antitrust Rules Already Split Safety From Output

The labs have spent years warning that a coordinated pause could draw antitrust fire. Amodei’s waiver ask was an admission of that risk. The complaint treats the public yes-chorus, and the private work behind it, as the agreement the waiver was supposed to legalize.

U.S. agencies have drawn this line before in a different industry. On April 10, 2014, the FTC and the Justice Department said properly designed sharing of cyber threat information is not likely to raise antitrust concerns. They also said that kind of sharing is very different from swapping current or future prices, output or business plans.

A later FTC note on competitor data exchanges made the sting explicit. If the sharing produces an agreement to fix prices or limit output, that agreement may be unlawful on its face and can even support a criminal case. Safety notes about a model’s failure modes sit closer to the 2014 cyber lane. A joint cap on how fast ChatGPT, Claude, Grok and Gemini get better sits closer to an output limit.

That is why the complaint’s carve-outs matter. Unilateral slowing is left alone. Joint slowing is the whole case. Amodei asked for a written hall pass for the second thing. Washington refused. The lawsuit now says they did the second thing anyway, in public, in about an hour.

Lehane’s Several Weeks of Quiet Lab Talks

The filing does not rest only on Saturday’s posts. It says the coordination started months earlier, and it names dates.

THE WEEK THE POSTS BECAME A PACT

1. July 2026: Representatives of Anthropic, OpenAI and Google form a working group on an industry standards body, the complaint says, and employees of frontier labs publish a July statement on pacing the frontier with 1,386 signatories, including Amodei, Anthropic co-founders Jared Kaplan and Jack Clark, OpenAI chief scientist Jakub Pachocki, OpenAI chief research officer Mark Chen, and Google DeepMind co-founder Shane Legg.
2. July 14, 2026: Hassabis publicly proposes a U.S.-led frontier-AI standards body modeled in part on the Financial Industry Regulatory Authority.
3. September 6, 2026: OpenAI publishes Pachocki’s essay “An Alien Mind,” which the complaint says listed coordination among frontier developers to slow future work as a main option.
4. September 10, 2026: The complaint cites a report that OpenAI had asked members of Congress whether an industry-wide slowdown could violate antitrust law.
5. September 12, 2026: Amodei publishes the essay; Musk, Altman and Hassabis endorse it within about an hour.
6. September 14, 2026: Altman says progress will go more slowly than it otherwise could, and that OpenAI will not wait for an antitrust exemption or new legislation before working with colleagues across the industry.
7. September 15, 2026: Chris Lehane, OpenAI’s head of global affairs, tells reporters the company has been operating “under the premise there is no need for an antitrust exemption,” and that it has already been working for several weeks with Anthropic, Google DeepMind and other labs on standards.
8. September 18, 2026: The four consumers file the Sherman Act complaint.

Lehane’s line is the one that will be read back in a motion to dismiss. OpenAI said it did not need a waiver. It also said the talks were already running. The plaintiffs say those talks, plus the public lockstep on September 12, are the conspiracy.

Anthropic has since posted measurements of how fast AI is advancing inside frontier labs, including how much of AI research is done by AI itself, and it noted those figures would shift if there were coordination on pacing. That is useful transparency. It is also a public trail of the very variable the complaint calls output.

The four companies had not commented by Friday night. The injunction the plaintiffs want would freeze the talks OpenAI’s policy chief said were already in motion.

Disclaimer: This article is news reporting and analysis of a civil complaint and related public statements. It is informational only and is not legal advice, investment advice or a prediction of how any court will rule. Readers who have a claim, a defense or a financial position tied to these companies should consult a licensed attorney and, where money is at stake, a qualified financial adviser before acting. The filing is an untested set of allegations, the labs have not answered in court, and figures and case status can change as the docket moves.