On July 23, 2026, a bipartisan bill, H.R. 9925, known as the FRONTIER Act and dedicated to federal oversight of the development of frontier artificial intelligence models, was introduced in the U.S. House of Representatives. The 74-page document ties regulation to a verifiable computational threshold of 10^26 FLOPs and monetary criteria, requires major laboratories to publish safety frameworks, undergo independent audits, and report critical incidents. The bill is currently in committee, so its significance is signaling: it is the most detailed available blueprint of how the United States might regulate the largest AI models.

What Happened

The bill's author is Republican Jay Obernolte from California, with co-sponsors including Lori Trahan, Erin Houchin, Scott Peters, Scott Franklin, and Suhas Subramanyam, representing both parties. The document has been referred to the Committee on Energy and Commerce and the Committee on Science, Space, and Technology. The FRONTIER Act defines a "frontier model" as a model whose training cost more than 10^26 operations, with the threshold including both fine-tuning and reinforcement learning stages. Developers are divided into two classes: "large" — with revenue over $50 million and AI development expenses of at least $1 billion over 36 months, and "very large" — with thresholds of $5 billion and $10 billion respectively. Large developers are required to publish a "frontier AI framework" document on risk and cybersecurity thresholds within one year of the law's signing, as well as undergo an annual audit without conflict of interest, in which the auditor is prohibited from being paid "for the result." "Very large" developers must engage IVO organizations licensed by the Department of Commerce with verification of internal model use. Section 8 grants the Department of Commerce emergency order powers, Section 9 preempts state laws, and GAO is tasked with studying the capacity of the future AI audit and verification market. The bill also introduces a quantified definition of catastrophic risk — more than 50 deaths or more than $1 billion in damage in a single incident — with a closed list of categories: chemical, biological, and nuclear weapons, cyberattacks without human involvement, and model escape from control. "Critical safety incidents" include the leak or theft of model weights, loss of control, and a situation where the model deceives its own developer's control. If the law is passed, the Department of Commerce must issue regulations within 180 days, and a report from the Deputy Secretary for AI Safety to Congress is scheduled from January 1, 2028.

Context

The tie of regulation to computation was not chosen by accident: training costs are auditable, whereas models' "dangerous capabilities" are poorly measured, so the 10^26 FLOPs threshold becomes a rare example of a verifiable trigger for legislation. The inclusion of fine-tuning and reinforcement learning stages in the threshold looks like an attempt to close loopholes through post-training. The FRONTIER Act is the first attempt in the 119th Congress to build comprehensive federal oversight specifically for frontier-AI, and it appears against the backdrop of the fact that independent replication of laboratory claims in current practice is almost absent. The requirement for "very large" developers to engage external IVOs effectively orders a reproducible independent assessment of safety claims, which institutionally is a step forward. The quantified definition of catastrophic risk with a closed list of categories serves as a ready-made task statement for risk quantification and sets a benchmark against which laboratories' eval infrastructure can be checked.

Why This Matters for the Industry

If passed, the regime would automatically cover the largest players — OpenAI, Anthropic, Google DeepMind, Meta, and xAI — already at the next generation of models, while startups remain outside the perimeter: the monetary thresholds deliberately concentrate the burden on the largest laboratories. For product and compliance teams, the document is already read as the most specific formalized blueprint: mandatory "frontier AI framework," annual audit without conflict of interest, reports on "critical safety incidents," and precise thresholds provide ready-made entities for compliance workflows and scope calculators. At the same time, the first federal market for AI audit and verification is being laid, where independent model assessment becomes a separate service, and eval infrastructure and computation accounting become a mandatory part of development. Even with low chances of passage, the text is worth analyzing as a benchmark: for the first time in the 119th Congress, there is a bipartisan document with precise thresholds and deadlines, showing where frontier-AI regulation is heading.

Why This Matters for Users

For readers, there are no direct legal changes at the moment: the bill is in committee, GovTrack estimates its chances of passage at 4 percent, and obligations only arise after the law is signed. If the FRONTIER Act is nevertheless passed, users will get more transparency: the safety frameworks of major developers will become public documents, independent IVOs will verify internal model use, and the Department of Commerce will have emergency intervention powers in critical situations. Today, the full text of the document is open on GovInfo, the status can be tracked on GovTrack or congress.gov, and the discussion on Hacker News is still purely symbolic — three points and one comment.

What Is Still Unknown / Limitations

The text fixes thresholds and subjects, but not measurement methodologies: it is not defined how to count and verify FLOPs, including distributed runs and computation estimation at reinforcement learning stages, and how to practically verify developers' claims about costs. The categories of "critical safety incidents" describe events that are currently detected extremely unreliably, so mandatory reporting without validated detection methods risks yielding either noise or a false sense of security. The fate of the document itself is uncertain: it is in committee, and there is no hearing or markup calendar in the sources. Expectations that the thresholds will become a common benchmark for discussion, and that demand for compliance platforms and amendment tracking dashboards will grow, remain interpretations, not facts. Outside the scope of the text are also questions of derivative model coverage, as well as the feasibility of loss-of-control detection — these are predictably set to become the focus of technical criticism if hearings begin.

Sources

Author

Look at AI, editorial team