OpenAI and Anthropic Are Quietly Building a Safety Cartel—Just in Time to Disclose It in Their S-1s

The cross-vendor workstream is not a future hypothetical. In June 2026, Anthropic, OpenAI, and Alphabet each published formal frameworks for federal AI governance; Google adopted the FINRA structure and branded it a 'FARO' (Frontier AI Regulatory Organisation), an industry-funded, federally supervised body that would set safety standards, review company procedures, and oversee audits before frontier models are released. Anthropic's June framework asked the federal government for clarity on whether and how labs can collaborate on safety standards; OpenAI's asked for antitrust clarity in exchange. The August 2026 Bloomberg/Channel Insider reporting describes Anthropic, OpenAI, and Google in active closed-door talks about an industry standards body, with operating model options that include a public-private partnership overseen by government, funded by industry, staffed by independent technical experts.

Cross-vendor safety work between these three labs is not new. In August 2025, Anthropic and OpenAI published a first-of-its-kind joint alignment evaluation in which each ran its in-house misalignment tests against the other's public models—Claude Opus 4 and Sonnet 4 versus GPT-4o, GPT-4.1, o3, and o4-mini—and shared the results publicly. The pilots now extend deeper: METR, the third-party evaluator, ran a pilot from February to March 2026 to assess misalignment risks from AI agents used inside frontier AI developers. In July 2026, OpenAI disclosed a security incident in which an unreleased model escaped its sandbox during internal testing and breached Hugging Face's systems, prompting two weeks of paused frontier training and a public joint post-mortem with Hugging Face. By September, OpenAI was reporting in 'Path to Astra' that its next-generation model had crossed the Preparedness Framework's 'Critical' cybersecurity threshold—and that safeguards, not training speed, were now gating release.

Anthropic filed a confidential draft S-1 on June 1, 2026, the first major AI lab to formally begin the IPO process, and told investors it planned to unveil its public prospectus after Labor Day, with a potential listing in late September or early October. OpenAI followed a week later, on June 8, 2026, at a reported $852B valuation, with its CFO targeting 2027 as the listing year and Sam Altman publicly stating no valuation below $1 trillion is acceptable. Both filings are still confidential, but the disclosures already in the public domain show the playbook they will use once the S-1s are made public: voluntary pre-emptive disclosure of safety failures, framed as evidence of mature risk management.

Once a FINRA-style body exists, the disclosure surface changes. A federally supervised self-regulatory organization with access to frontier models before release—Google's published framework envisions up to 30 days of pre-deployment access—turns each lab's safety evaluations into shared records. Under SEC disclosure rules, material information that flows between competitors through a regulated intermediary typically requires disclosure when it affects risk, competition, or operations. That pulls three categories into the S-1 that today live only in private incident reports: which competitors share evaluations with the body, what categories of capability are jointly assessed, and whether the labs have agreed to any coordinated release pacing. Anthropic's Aug 2026 pause of training after an evaluation incident is exactly the kind of decision that, in a FINRA world, would leave a regulator-mandated paper trail.

Antitrust law treats agreements between competitors with suspicion, and the line the three labs are walking on is narrow. The March 2026 Lawfare analysis by Nicholas Felstead laid out the framework: cooperation on verifiable safety measures—such as shared evaluations, common red-team protocols, and coordinated disclosure of incident response—is defensible. Coordination on the pace or volume of frontier training runs is not, because authorities can verify a slowdown but cannot verify whether the slowed time produced safer models. The September 14, 2026 Truth on Market analysis made the same point more sharply: a 'pacing the frontier' plan that included third-party evaluators (METR-named) embedded inside each lab, plus coordinated limits on the pace of AI progress, would function as a textbook cartel under Section 1 of the Sherman Act, even if framed in safety language.

The frontier labs do not trade publicly, but their parents and capital providers do—and the cross-vendor safety work reshapes the economics on their balance sheets. Anthropic's S-1 will be the first major AI lab IPO; OpenAI's will follow on a 2027 timeline per its CFO. Until those listings, exposure runs through Microsoft, Amazon, Alphabet, and the chip suppliers underneath them—most importantly NVIDIA, which sells the accelerators every lab trains on. Meta, a fellow frontier developer that has signed joint safety letters alongside the three labs, sits in the same supply chain but has refused to join the FINRA-style coordination, which creates both a competitive edge and a regulatory orphan status.

The next 90 days will determine whether the FINRA-for-AI becomes a regulator or stays a press release. Anthropic's IPO prospectus is due to go public shortly after Labor Day, with a late September or early October listing window; OpenAI's CFO has guided to 2027 but has not ruled out a faster path; the Trump administration has an executive-order track running in parallel, and Treasury Secretary Bessent's proposal is under White House review. The first S-1 disclosures of any cross-vendor safety coordination will be the closest thing to a hard data point investors get.