Class action complaint, Buist v. Anthropic, PBC (N.D. Cal., No. 3:26-cv-10693, filed September 18, 2026)

The complaint in Buist v. Anthropic, PBC, No. 3:26-cv-10693, was filed on September 18, 2026 in the United States District Court for the Northern District of California, San Francisco Division. It runs to 29 pages and 152 numbered paragraphs and is brought by four individual subscribers to paid AI services, for themselves and a proposed nationwide class, against Anthropic, PBC, OpenAI OpCo, LLC, SpaceXAI LLC and Google LLC. It alleges that the four companies agreed to slow the pace at which each improves the products it sells to consumers, and that the agreement is a horizontal restraint on output and quality prohibited by Section 1 of the Sherman Act, 15 U.S.C. § 1. The case-tracking page is Buist v. Anthropic, PBC; this page summarises the pleading itself.

The complaint's central theory is that an agreement among competitors can be formed entirely in public. It pleads Dario Amodei's September 12, 2026 essay "We Must Pace the Frontier" as the offer, and statements the same day by Elon Musk, Sam Altman and Demis Hassabis as acceptances: in its words, the agreement "was proposed in public, accepted in public, and confirmed in public" (¶ 2). Counsel are Nicholas C. Rowley (lead counsel), Andrew T. Tutt, R. Stanton Jones and Jakob Z. Norman of Trial Lawyers for Justice; Jones and Norman are listed with pro hac vice applications forthcoming. Tutt signed the complaint.

Parties

Plaintiffs (¶¶ 16–21). Charles Buist and Nick Spetsas, residents of Florida, and Cheyenne Hunt, a resident of California, each bought paid individual consumer subscriptions to Claude, ChatGPT, Grok and Gemini during the class period. Christine Bullock, a resident of California, subscribes to Claude only. All four continue to subscribe and intend to keep doing so, which the complaint relies on for standing to seek an injunction. Each is alleged to have subscribed to obtain the defendant's most capable models and continual improvements, expecting each defendant to decide independently how aggressively to develop its product.

Defendants (¶¶ 22–32). The complaint identifies each defendant's consumer subscription products and the individuals whose statements it attributes to that defendant:

The complaint alleges that unnamed persons, "including affiliates of Defendants and other frontier-AI developers," participated in or knew of the conduct, reserves the right to add them, and pleads each defendant as a co-conspirator liable for the others' acts (¶¶ 34–35).

Framing and stated limits

The introduction sets the plaintiffs' position on safety before the factual allegations. It states that the plaintiffs "take AI safety seriously" and believe AI "should have guardrails with ethical considerations and direct liability for any breach being paramount," but that those guardrails "should not be decided by the Defendants, but by we the people, through state and federal regulation, and through the voices of juries" (¶ 7). It cites AI's use in protein-structure prediction, cancer imaging, materials discovery and mathematics as gains produced by "the competitive race that Defendants' agreement now threatens to slow" (¶ 8).

Two analogies carry the argument that safety is a unilateral matter. Automobile manufacturers do not need a horizontal agreement to keep passenger cars from reaching 300 miles per hour, because engineering constraints, internal standards, product liability, regulation by the National Highway Traffic Safety Administration and competitive self-interest each supply the limit independently (¶ 9). On environmental impact, the complaint cites the International Energy Agency's projection that global data-centre electricity demand will more than double from 2024 levels by 2030, and argues that efficiency is itself a field of competition that an agreement to slow capability would not advance (¶ 10). It concludes that "the antitrust laws do not permit competitors to decide among themselves that competition is too dangerous," and that whether frontier AI should develop more slowly "is a question for each company acting alone, or for Congress and the agencies" (¶ 12).

The complaint disclaims any challenge to a defendant's unilateral decisions on safety, testing, environmental impact or the pace of its own development, and to any defendant asking Congress for regulation or an antitrust exemption (¶ 13). A separate section states that the plaintiffs do not seek liability for advocacy to Congress, the White House or any agency, "including advocacy for AI regulation, for a standards body created by statute, or for an antitrust exemption"; the petitioning statements are pleaded "only as evidence of Defendants' knowledge, intent, and understanding" of the competitive problem and of the antitrust risk they accepted (¶¶ 88–90).

Factual allegations

Competition through the pace of improvement

The complaint alleges that ChatGPT, Claude, Grok and Gemini are close substitutes, that the defendants compete principally on capability and secondarily on usage limits, context windows, memory, integrations and speed of access to new models, and that paid tiers monetise those differences directly (¶¶ 43–45). Each major release causes consumers to subscribe, cancel or switch, and each defendant monitors and responds to the others' releases, benchmarks and pricing (¶ 46). It cites OpenAI researchers' description of using AI to accelerate AI research, and Anthropic's description of recursive self-improvement as a driver of accelerating capability, for the proposition that an agreement about today's pace shapes every later generation of products (¶ 48).

Events before September 12

The complaint pleads a sequence of statements it treats as recognising that no firm would slow alone and building the means to coordinate:

• The July 2026 statement. It alleges that "senior executives of Anthropic, OpenAI, and Google" published a statement titled Pacing the Frontier, with signatories including Amodei, Kaplan, Pachocki, OpenAI Chief Research Officer Mark Chen, Legg and "numerous other employees," and quotes its acknowledgment that each company is under "intense competitive pressure not to unilaterally slow" (Pacing the Frontier (statement from employees of frontier AI companies, July 2026); ¶¶ 49–50). It alleges that Altman told reporters in July that OpenAI had helped shape the statement's language and that he had discussed the "need" to pace AI development with White House officials (¶ 51).
• Hassabis's standards-body proposal. On July 14, 2026 Hassabis proposed a U.S.-led standards body for frontier AI modelled in part on the Financial Industry Regulatory Authority, to which laboratories would initially submit models voluntarily for pre-release review and which could later impose and escalate common requirements (A Framework for Frontier AI and the Dawning of a New Age (Hassabis, July 2026); ¶ 53). The complaint states that contemporaneous reporting described the proposal as a mechanism that could coordinate an industry-wide slowdown, and characterises such a body as a standing forum for exchanging information and coordinating responses to capability thresholds (¶¶ 53–54).
• The working group. It alleges that representatives of Anthropic, OpenAI and Google below the CEO level formed a working group in July 2026 that met regularly on creating an industry standards body; that The Information reported on September 13 that its meetings were continuing and that it had met the preceding week; and that OpenAI confirmed on September 15 that it had been discussing AI safety with Anthropic and Google DeepMind for several weeks (¶ 55).
• Altman's town hall. Altman is alleged to have told an OpenAI companywide town hall that the major laboratories might need to create a standards body themselves if government did not, and, according to published reports, that OpenAI might slow frontier development, "potentially in conjunction with other laboratories" (¶ 56).
• Pachocki's essay. On September 6, 2026 OpenAI published Pachocki's "An Alien Mind," which the complaint quotes as naming "coordinating to slow down future development" as one of the principal options and as expecting voluntary slowdowns to become common until shared safety thresholds exist (An Alien Mind (Jakub Pachocki, OpenAI, September 2026); ¶ 57).
• OpenAI's inquiry to Congress. WIRED reported on September 10 that OpenAI had asked members of Congress whether coordinating an industry-wide slowdown could violate the antitrust laws; the complaint alleges OpenAI did not substantively deny the report, and reads the inquiry as OpenAI's understanding that the contemplated conduct was coordination among competitors (¶ 58).
• The Fortune interview. On September 11, asked by Fortune why leading executives including Amodei and Musk did not develop a common plan, Altman answered "I think that will happen" and declined to "pre-announce private discussions" that should later be shared "as a group" (¶¶ 59–60).

The essay as offer

The complaint quotes the essay's thesis, "We must slow the pace at which we improve the capabilities of AI models," and describes its three steps (¶¶ 61–68). The first, embedding third-party evaluators with employee-like access, Anthropic adopted unilaterally. The second, which the complaint calls "Democratic Coordination," the essay distinguished as requiring "industry-wide coordination" on common safety standards and "limits on the rate of unchecked AI progress." The complaint reads the essay as proposing limits on training compute, on the nature of training runs and on internal use of AI to improve AI, and capability "checkpoints" at which further advancement would depend on agreed certifications; it reads the embedded evaluators as the means of enforcement, making "verifiable pacing" practical by letting outsiders confirm that each participant was slowing as agreed (¶¶ 65–66). It quotes the essay's statement that coordination would let developers slow "without sacrificing commercial advantage," and characterises that as "the economic function of every output-restricting cartel" (¶ 67). It notes the essay's call for companies to "voluntarily work together" alongside legislation, and its statement that the government would need to issue a narrow waiver for certain safety conversations; the complaint states that no waiver, statutory exemption or agency compulsion exists (¶¶ 68–69).

Paragraph 70 maps the essay onto the elements of a horizontal agreement: the parties (the frontier laboratories), the term (a slower rate of advancement than each would choose alone), the mechanisms (shared limits on compute, training runs and self-improvement, and capability checkpoints) and the means of "policing defection" (verifiable pacing through embedded evaluators).

Acceptances and confirmation

The complaint alleges that within about an hour Musk quote-posted the proposal with "Dario is right" (¶ 71); that later that day Altman wrote "I agree with Dario that we need to pace the frontier," said pacing had been a primary subject of discussion inside OpenAI in recent weeks, stated that OpenAI "will do the same" in giving independent evaluators employee-like access, and said OpenAI would have more to share soon (¶ 72); and that Hassabis said the essay pointed toward "the right path forward" and that "the direction is correct," while noting that details needed further work, and tied it to his proposed "industry-wide standards body for frontier AI" (¶ 73). It pleads that by the end of September 12 a senior executive or controlling person of each defendant had signalled assent, in sequence and with knowledge of the others, and that "a public offer followed by public acceptances forms an agreement no less than" private communications (¶¶ 74–75).

On September 14 Altman is alleged to have said that OpenAI looked forward to "collaborating with our colleagues across the industry" on common safety approaches, that AI progress "should be slower than it otherwise could be," and that OpenAI did not intend to wait for an antitrust exemption or legislation (¶¶ 76–78). On September 15 Lehane is alleged to have confirmed that OpenAI had been working with Anthropic and Google DeepMind on these issues for several weeks, before the essay, and that the firms could work together without waiting for a waiver; the complaint reads this as establishing that the September 12 statements were "the public face of an existing course of dealing" (¶¶ 79–80). It alleges the coordination continues through Amodei's advocacy, Altman's anticipated announcements, the working group and the proposed standards body (¶ 81).

Each defendant's participation and the plus factors

Paragraphs 82–85 set out each defendant's alleged role. Anthropic proposed the agreement, defined its terms, supplied its enforcement mechanism and joined the working group from July. OpenAI joined the working group, helped draft the July statement, explored coordination with employees and the White House, sought guidance from Congress, published Pachocki's essay, accepted, stated the object and confirmed prior collaboration. Google joined the working group, proposed the standards body through Hassabis, had a Google DeepMind co-founder sign the July statement, and endorsed the essay while directing implementation toward its own proposed body. SpaceXAI, through Musk, endorsed the proposal within an hour, "communicating to its three principal rivals that SpaceXAI would not exploit their restraint by racing ahead"; the complaint adds that the Fortune question Altman answered on September 11 named Musk as a participant in the contemplated plan.

To distinguish agreement from parallel conduct, the complaint pleads direct evidence (the written proposal, express statements of agreement, and admissions of weeks of private collaboration) and six circumstantial factors: motive; action against unilateral self-interest, since no firm would announce slower improvement absent assurance that rivals would match; interfirm communications from July through September; market structure; consciousness of antitrust risk, shown by the congressional inquiry, the essay's waiver request and Altman's statement that OpenAI would proceed without one; and pretext, on the ground that every stated safety objective can be pursued unilaterally or through regulation (¶¶ 86–87).

Market power and effects

The complaint pleads first that the agreement is a naked horizontal restraint unlawful without proof of a relevant market (¶ 95). In the alternative it defines a "Paid Frontier AI Assistant Subscription Market" of paid consumer subscriptions to general-purpose frontier generative-AI assistants in the United States, principally the paid tiers of ChatGPT, Claude, Grok and Gemini (¶¶ 96, 101). Search engines, productivity software, narrow machine-learning tools and human professional services are pleaded as outside the market under a test framed as a small but significant non-transitory decrease in quality or rate of improvement, or increase in price (¶ 98). The defendants' free tiers are excluded because of lower limits and restricted model access, and because the agreement restrains free and paid tiers alike (¶ 99). On information and belief the defendants account for at least 80% of U.S. paid subscriptions in the market (¶ 102). An innovation market for the development of general-purpose frontier generative-AI models is pleaded in the further alternative (¶ 100). Entry barriers include compute, specialised semiconductors, electricity, data-centre capacity, talent and the ability to fund successive model generations (¶ 103).

The alleged effects are reduced incentive to invest in capability, begin or expand training runs and respond to rivals; delayed and fewer improvements; and a higher quality-adjusted price, because subscribers pay the same or higher nominal prices for products that improve more slowly (¶¶ 105–109). The complaint acknowledges that, because the agreement is recent and development cycles run months, "its full effect on released products has not yet manifested," but alleges on information and belief that it has already altered investment, training and release decisions (¶ 110). It argues that the agreement has no cognisable procompetitive justification, that a contention that competition in frontier AI is itself unsafe is not one cognisable under the Sherman Act because "that judgment belongs to Congress," and that the agreement is not ancillary to any integration of productive assets or joint venture (¶¶ 111–114).

Antitrust injury is pleaded as an overcharge: subscribers pay prices exceeding the competitive value of what they receive, measurable class-wide through development plans, compute and training decisions, benchmark performance, prices and switching data (¶¶ 115–118). Plaintiffs are pleaded as direct purchasers and the most efficient enforcers (¶ 120).

Class allegations

The proposed nationwide class, under Rules 23(a), 23(b)(2) and 23(b)(3), is all persons in the United States who from September 12, 2026 until the effects of the conduct cease bought directly from a defendant a paid individual consumer subscription providing premium access to ChatGPT, Claude, Grok or Gemini (¶ 123). Four defendant-specific subclasses are proposed, one per defendant (¶ 124). The plaintiffs reserve the right to modify the definitions, including on evidence that the agreement formed before September 12 (¶ 125). Excluded are the defendants and their affiliates, officers and directors, governmental entities, and the court and its staff (¶ 126). Common questions listed include whether the defendants communicated privately before or after the September 12 statements, and whether the restraint is unlawful per se, under abbreviated analysis or under the rule of reason (¶ 128).

Claims and relief

The injunction sought would bar each defendant from entering, maintaining, enforcing or monitoring any agreement with a horizontal competitor concerning (a) the rate at which competing AI products are developed, improved, trained or released; (b) limits on training compute or on the nature, frequency or timing of training runs adopted under a horizontal pacing arrangement; (c) limits on the use of AI to develop improved AI, where imposed by agreement among competitors; (d) coordinated delays in releasing models or features; (e) capability checkpoints operating as horizontally agreed restraints, "as distinct from lawful governmental requirements or independent safety measures"; and (f) exchange of competitively sensitive information to police any such restraint (¶ 150). It would not bar independent safety measures, independent slowing or suspension of development, independent retention of evaluators, lawful safety research, compliance with government requirements, petitioning, or "legitimate standard-setting activities that do not restrain independent competition over the pace of development" (¶ 151).

The prayer asks for class certification, a declaration of a Sherman Act violation, trebled damages, preliminary and permanent injunctive relief, pre- and post-judgment interest, and fees and costs (A–G). The plaintiffs demand a jury trial.

Jurisdiction and venue

Subject-matter jurisdiction is pleaded under 28 U.S.C. §§ 1331 and 1337(a) and 15 U.S.C. §§ 15 and 26, and under the Class Action Fairness Act on the allegation of more than 100 class members, minimal diversity and more than $5,000,000 in controversy (¶¶ 37–38). Assignment to the San Francisco Division rests on Anthropic's and OpenAI's principal places of business there (¶ 40). Personal jurisdiction and venue rely on Anthropic, OpenAI and Google being at home in the district, SpaceXAI's sales there, and nationwide service under Section 12 of the Clayton Act (¶¶ 41–42).

Allegations compared with other accounts

The complaint's account of September 12–15 matches the record on We Must Pace the Frontier (Amodei, September 2026) and Coordinated slowdown proposals compared on timing and wording: Musk's "Dario is right" about an hour after publication, Altman's post, Hassabis's post that evening, Altman's September 14 statement that OpenAI did not need an exemption to begin, and Lehane's September 15 account. Several of its characterisations differ from how other sources describe the same material, and are allegations rather than established facts:

• The July statement. The complaint describes Pacing the Frontier as published by "senior executives of Anthropic, OpenAI, and Google." Pacing the Frontier (statement from employees of frontier AI companies, July 2026) records it as an open statement signed by employees of frontier AI companies in a personal capacity, including employees of Meta and Thinking Machines, and as requesting U.S. government support for an international effort to build pacing tools rather than an industry arrangement. Its sentence reads that each "company—and country—is under intense competitive pressure not to unilaterally slow"; the complaint quotes it without "and country." The complaint names Shane Legg as a signatory and pleads that a Google DeepMind co-founder signed; the captured list (the first 30 of 1,224 signatories) does not include Legg, so his signature is not independently confirmed, while Google DeepMind's Jasjeet Sekhon and Google's Anca Dragan do appear.
• Input limits. The complaint reads the essay as proposing limits on training compute, training runs and internal AI-driven AI development. The essay page records Amodei as expressing most enthusiasm for pacing keyed to capability and observed safety, and raising input-based pacing while noting that such measures "may be more 'gameable'."
• The second step. The essay frames coordination among frontier companies "in democracies" and prefers regulation reaching all U.S. frontier companies as "the most effective method of pacing"; the complaint's account centres the voluntary, industry-wide element.
• Lehane's statement. Other coverage records Lehane as saying he did not see the need for an antitrust waiver for the three companies to talk about safety (Source: bloomberg.com); the complaint pleads the statement more broadly, as confirming that the firms could "work together" on pacing without a waiver.
• Uncorroborated items. Altman's July remarks that OpenAI helped shape the statement's language, the Fortune interview of September 11, and the characterisation of the working group as meeting "regularly since July" rest on the complaint's own account and the reporting it cites; other sources record the three companies' standards-body discussions as predating the essay without dating their start.

Key claims

Provenance

The raw file is CourtListener's plain-text extraction of the RECAP copy of docket entry 1 (29 pages, filed and entered September 18, 2026), captured September 21, 2026 with pleading-paper line numbers and page headers retained. No exhibits were filed with the complaint; the certificate of interested entities and proposed summons were not captured. Authenticity was checked against the CourtListener docket (party list, nature of suit 410 Anti-Trust, cause 15:1, filer Andrew Tutt) and the Associated Press report of the filing (Source: courtlistener.com; broadbandbreakfast.com).

Relationships

• supports: Buist v. Anthropic, PBC — the pleading the case page tracks.
• depends-on: We Must Pace the Frontier (Amodei, September 2026) — pleaded as the offer.
• depends-on: Pacing the Frontier (statement from employees of frontier AI companies, July 2026) — cited for the collective-action problem.
• depends-on: A Framework for Frontier AI and the Dawning of a New Age (Hassabis, July 2026) — pleaded as the coordination vehicle.
• depends-on: An Alien Mind (Jakub Pachocki, OpenAI, September 2026) — cited for "coordinating to slow down future development."
• instance-of: AI Antitrust — private Sherman Act enforcement directed at proposed safety coordination.
• related: Coordinated slowdown proposals compared — the proposals and exemption debate the complaint responds to.
• related: Collaboration on Adversarial Threats and Security Risks Act (S. 5105) — the pending antitrust exemption for delay agreements among developers; the complaint states no exemption has been enacted.
• litigates: Anthropic, OpenAI, xAI, Google DeepMind — defendants (Google LLC is the named entity).
• related: Dario Amodei, Sam Altman, Elon Musk, Demis Hassabis, Chris Lehane, Jakub Pachocki, Jared Kaplan, Shane Legg — individuals whose statements the complaint attributes to the defendants.