Anthropic S-1: AI safety, competition law and the China problem
In short: In the Anthropic S-1, its US listing document filed on 26 October 2026, Anthropic suggests that safety standards agreed with the US government and other frontier AI developers, including on the pace of development, may warrant an antitrust safe harbour (a legal exemption). US, EU, UK and Chinese competition law each treat an agreement between competitors to limit technical development as a likely restriction, and there are no existing exemptions permitting co-operation on the grounds of ensuring AI safety.
A private agreement between rival AI developers to slow the development or release of more capable models would be at serious risk of infringing competition law in the US, the EU, the UK and China, and we have found no exemption for AI safety. Anthropic has proposed that safety standards agreed with the US government and other developers may warrant a narrow antitrust exemption, and it is defending a US class action alleging that it has already agreed to slow down. An agreement among US developers would bind no one outside it, and Anthropic’s own filing records the risk that developers in less regulated countries may move faster.
Safety cooperation and antitrust in the Anthropic S-1
Anthropic, PBC filed its registration statement on Form S-1 with the US Securities and Exchange Commission on 26 October 2026. In its risk factors (page 98), Anthropic states that it has proposed working with the US government and other frontier developers on safety standards, “including with respect to the pace of capability development”, and has suggested that this collaboration may warrant “a narrow exemption or safe harbor” from antitrust law. It records public opposition from political leaders, some of whom have called such efforts anticompetitive whatever their purpose. It also discloses that in September 2026 claimants filed a putative (proposed) class action in the US District Court for the Northern District of California against Anthropic and other frontier developers, alleging an agreement to slow development contrary to section 1 of the Sherman Act and seeking treble damages (three times the loss) and an injunction. Anthropic disputes the claims.
Anthropic also warns (page 70) that “increasingly restrictive AI regulation in the United States and other western jurisdictions could enable competitors operating in less regulated environments to innovate, scale, and deploy AI technologies more rapidly”, and says that strategic competition between the US and China may lead to divergent regulation.
The United States: partial protection for joint research and standards bodies
US law protects some forms of cooperation between competitors, and on our reading that protection is unlikely to extend to an agreement on pace. Under the National Cooperative Research and Production Act (15 USC 4301 to 4306), a joint venture for research, development, testing or production, and a standards development organisation, is judged under the rule of reason (an assessment of actual effects) rather than treated as illegal per se (section 4302). Where the venture or standards body has notified the Department of Justice and the Federal Trade Commission, a successful claimant recovers actual rather than treble damages (sections 4303 and 4305). Neither protection is likely to cover a pace agreement. The definition of a joint venture excludes agreements restricting the parties’ own provision of products, or their other research and development, unless reasonably required for the venture (section 4301(b)(2) and (3)), and the protection for standards bodies covers the organisation rather than the developers that take part in it (sections 4301(a)(8) and 4302). Losing that protection does not make an agreement unlawful in itself: a restraint ancillary to a genuine collaboration is generally judged on its effects, and joint lobbying of government for regulation is treated differently again.
The agencies withdrew their 2000 guidelines on collaborations among competitors in December 2024, and in February 2026 asked for comments on whether to issue new guidance. We have not found replacement guidance, so no current agency guidance addresses what developers may agree.
The European Union: safety duties set by legislation
Article 101(1)(b) of the Treaty on the Functioning of the European Union (TFEU) prohibits agreements that “limit or control production, markets, technical development, or investment”, and Article 101(3) exempts an agreement that meets four conditions on efficiency, consumer benefit, indispensability and residual competition. In its 2023 Horizontal Guidelines the Commission treats a sustainability standard as unlikely to restrict competition where it meets six conditions: an open and transparent process, no obligation on non-participants, freedom for participants to go further, limited exchange of sensitive information, open access, and either no significant price rise or a combined market share below 20% (section 9.3.2.4). In the same Guidelines the Commission treats an agreement between competitors to limit technological development to the legal minimum as a restriction by object, one treated as harmful without proof of its effects (section 9.3.2.3), citing its Car Emissions decision of 8 July 2021, in which car makers agreed not to improve emission cleaning beyond what the law required. An agreement to slow AI development would fall outside that safe harbour, which requires participants to stay free to go further and does not extend to agreements limiting output (footnote 378). Its aim would be to raise safety rather than to hold it at a legal minimum, but it would share with Car Emissions an agreed limit on technical development.
EU legislation imposes safety duties on the most capable models. Providers of general-purpose AI models with systemic risk must evaluate their models, assess and mitigate systemic risks, report serious incidents to the Commission’s AI Office and protect their models’ cybersecurity (Article 55 of the EU AI Act). The AI Office facilitates codes of practice under Article 56, and a provider may rely on a code to show compliance until a harmonised standard is published, or show compliance by other adequate means (Article 55(2)). Anthropic, Google, OpenAI and others have signed the General-Purpose AI Code of Practice, published on 10 July 2025; xAI has signed its safety and security chapter only, and no Chinese developer appears on the Commission’s list. The code was drawn up through a process convened by the AI Office, and developers that sign it apply a common standard without having to agree one among themselves. The Act does not regulate the pace of development.
The United Kingdom: an analysis close to the EU’s
UK competition law applies provisions closely modelled on Article 101 TFEU. Section 2(2)(b) of the Competition Act 1998 uses the words of Article 101(1)(b), section 9 applies the conditions of Article 101(3), and section 60A requires the Competition and Markets Authority (CMA) and the courts to seek consistency with EU principles and case law from before 31 December 2020, unless they consider it appropriate to depart. The 2023 EU Guidelines and the Car Emissions decision came later and do not bind the CMA, but the CMA’s own 2023 guidance on horizontal agreements takes a similar approach: standardisation agreements with open participation, a transparent process, no obligation to comply and access on fair, reasonable and non-discriminatory terms will normally not restrict competition (paragraph 9.17). The UK has two tools of its own: a ministerial power to exclude agreements from the prohibition on exceptional and compelling grounds of public policy (Schedule 3, paragraph 7), and the CMA’s Green Agreements Guidance, which relaxes, for climate change agreements, the condition that consumers receive a fair share of the benefit. We are not aware of either being used for AI.
China and the problem of global alignment
China’s Anti-Monopoly Law, as amended with effect from 1 August 2022, prohibits agreements between competitors that restrict the development of new technologies or products (Article 17(4); only the Chinese text is authoritative, and English versions are translations). Article 20 disapplies the prohibition where the parties prove that the agreement is for technical improvement or new product development, improves quality or efficiency by unifying product specifications or standards, or serves public interests such as energy conservation and environmental protection, and also prove that it will not substantially restrict competition and lets consumers share the benefit. Public-facing generative AI services are regulated by the state: a provider whose service has public opinion attributes or social mobilisation capacity must undergo a security assessment and file its algorithms with the Cyberspace Administration of China (Article 17 of the Interim Measures for the Management of Generative AI Services, in force since 15 August 2023).
Zhipu.ai, MiniMax and 01.ai have signed the Frontier AI Safety Commitments made at the AI Seoul Summit in May 2024, under which each signatory sets thresholds for severe risk and commits, in the extreme, “not to develop or deploy a model or system at all” if it cannot keep risks below them. Those are voluntary commitments made by each company separately, and are not agreements between competitors on pace. No US antitrust exemption would bind a developer in China, and the US government acts on Chinese AI development through export controls and other national security measures rather than competition law. Those controls also apply to US developers: Anthropic discloses that in June 2026 the Department of Commerce imposed worldwide export restrictions on two of its models, and lifted them after Anthropic updated its safeguards in coordination with the government.
| Jurisdiction | Rule an agreement on pace would engage | Routes to lawful cooperation | Who sets frontier AI safety standards |
|---|---|---|---|
| United States | Sherman Act, section 1 | Rule of reason and actual damages for notified joint ventures and standards bodies (15 USC 4302 to 4305), which are unlikely to cover a pace agreement | Developers, with a voluntary federal framework for pre-deployment engagement |
| European Union | Article 101(1)(b) TFEU | Article 101(3); sustainability standards safe harbour as a model, though a pace agreement would fall outside it | The EU AI Act imposes safety duties, with a code of practice convened by the AI Office; it does not regulate pace |
| United Kingdom | Competition Act 1998, section 2(2)(b) | Section 9; public policy exclusion order under Schedule 3, paragraph 7 | Developers, with voluntary testing by the AI Security Institute |
| China | Anti-Monopoly Law, Article 17(4) | Article 20 exemptions for technical improvement and public interest | The state, through security assessment and algorithm filing for public services |
Viewpoint
In each of these jurisdictions, developers that agree to slow down would probably have to justify the agreement, and we have found no exemption built for it. In the EU, legislation imposes safety duties on the most capable models, and China regulates public generative AI services through security assessment and filing. Neither sets a pace of development, but both reduce the need for developers to agree safety standards among themselves. The US has not taken that step, and Anthropic has suggested a safe harbour instead.
Safety cooperation is weaker if any leading developer stays outside it, and no exemption granted in the US, the EU or the UK would cover a developer in China. The Seoul commitments are the only developer safety commitments we know of that both US and Chinese frontier developers have signed, and they are voluntary. On current law, each country sets the trade-off between pace and safety for itself, through regulation and export controls.
Frequently asked questions
What does the Anthropic S-1 say about antitrust and AI safety?
Anthropic states that it has proposed working with the US government and other frontier developers on safety standards, including on the pace of capability development, and has suggested that this collaboration may warrant a narrow exemption or safe harbour from antitrust law. It also discloses a putative US class action, filed in September 2026, alleging that frontier developers agreed to slow model development in breach of section 1 of the Sherman Act.
Can competing AI developers agree to slow development under EU and UK competition law?
An agreement between competitors to limit technical development is likely to fall within Article 101(1)(b) TFEU and section 2(2)(b) of the Competition Act 1998, and the Commission treats an agreement to limit technological development to the legal minimum as a restriction by object. On our analysis, such an agreement would very likely need to meet the four exemption conditions or, in the UK, be covered by a ministerial order excluding it on public policy grounds. A block exemption for joint research would rarely help, because limits on output, and on the parties’ freedom to carry out research outside the project or after it ends, are hardcore restrictions under the UK order (article 10 of SI 2022/1271).
How does Chinese competition law treat an agreement on AI development?
Article 17(4) of China’s Anti-Monopoly Law prohibits agreements between competitors that restrict the development of new technologies or products. Article 20 exempts an agreement that the parties prove is for technical improvement, new product development, common standards or a public interest, provided it does not substantially restrict competition and consumers share the benefit. Public generative AI services in China must pass state security assessment and algorithm filing.
This post was researched and drafted with the help of Claude, Anthropic’s AI model. Rob Bratby chose the topic and thesis, and reviewed and edited the final post. Rob takes full responsibility for its content.
For advice on EU and UK competition law and AI cooperation, including the AI governance questions that arise in safety forums, contact Rob Bratby at Bratby Law.
