What happened
Crypto Briefing reports that Anthropic's draft federal bill is facing specific criticism regarding its definition of 'frontier' models, which relies on a 10^25 FLOPs training threshold. Critics argue this figure is obsolete, noting that 33 models from 12 developers had already exceeded it by June 2025. The report details the bill's two-part test for 'covered developers,' requiring both the compute threshold and significant financial metrics, while contrasting it with California's SB 53 and the EU .
Crypto Briefing reports that the central point of contention in Anthropic's draft federal bill is the 10^25 FLOPs threshold used to define 'frontier' models. The article states that critics argue this figure reflects an earlier era of AI and now captures most enterprise-grade models, citing data that at least 33 models from 12 developers had already passed this mark by June 2025.
The proposal, which builds on Anthropic's Advanced AI Framework introduced in June 2026, defines a 'covered developer' through a two-part test. In addition to the 10^25 FLOPs requirement, a company must either generate more than $500 million in annual AI-derived revenue or spend over $1 billion in annual AI research and development. Companies meeting these criteria would face transparency obligations, though the specific list of obligations is not detailed in the source text.
The report contrasts Anthropic's threshold with other regulatory frameworks. California’s SB 53 sets its trigger at 10^26 FLOPs, which is ten times higher than Anthropic's figure. The EU uses 10^25 FLOPs as a rebuttable presumption of systemic risk, whereas Anthropic's draft functions as a hard compliance trigger. The article also notes that some critics view the draft as narrower and potentially weaker than the FRONTIER Act (H.R. 9925) regarding rulemaking authority.
A significant portion of the criticism focuses on the 'incumbent-advantage' argument. The report explains that compliance requirements such as safety frameworks and independent evaluations are costly and resource-intensive. Large labs already maintain teams for these tasks, while smaller players generally do not, meaning the bill's structure could disproportionately burden smaller developers despite the financial bars included in the two-part test.
Source details: cryptobriefing.com ↗
Why it matters
The debate over the 10^25 FLOP threshold highlights a structural flaw in using static metrics for dynamic AI regulation. If the threshold is too low, it captures most enterprise models, creating compliance burdens that favor large labs with existing safety teams over smaller players. This 'incumbent-advantage' critique, alongside the narrow scope of government intervention limited to catastrophic risks, shapes the political viability of the bill and the future regulatory landscape for AI developers.
The static nature of the 10^25 FLOP threshold exposes a structural problem in -based regulation: fixed numbers age quickly as hardware improves and training runs grow. The fact that 33 models already exceeded this limit by mid-2025 suggests that a static line drawn today may cover dozens of models tomorrow, potentially rendering the 'frontier' distinction meaningless or overly broad.
The bill's limitation on government power is also a critical policy development. The draft restricts the Commerce Department to restricting deployment only when catastrophic risks are foreseeable, defined as incidents causing damages exceeding $1 billion or the loss of more than 50 lives. Other agencies are barred from halting deployment altogether. This trade-off provides developers with predictability but leaves regulators with little room to act on harms that fall short of catastrophe, a point of contention when compared to other proposals.
The 'incumbent-advantage' critique has practical implications for market competition. If compliance costs are high and only large labs can easily absorb them, the bill could inadvertently create barriers to entry for smaller AI companies. This dynamic could influence how lawmakers amend the bill to ensure it does not stifle innovation from smaller players while still addressing safety concerns for large-scale models.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
Why can ethical evaluation not be reduced to one model score?
What to watch next
Watch for legislative amendments to the FLOP threshold or the addition of dynamic metrics, as well as reactions from smaller AI developers and the outcome of comparisons with the FRONTIER Act (H.R. 9925) in congressional hearings.
Legislators may seek to amend the 10^25 FLOP threshold to a higher number or introduce a dynamic metric that adjusts over time to prevent the 'frontier' category from becoming too broad. The reaction to the June 2025 data showing 33 models already exceeding the limit will likely drive these discussions.
The comparison with the FRONTIER Act (H.R. 9925) will be crucial. Critics who view Anthropic's bill as weaker on rulemaking authority may push for provisions from the FRONTIER Act to be incorporated, or vice versa, depending on which bill gains more traction in Congress.
Smaller AI developers and industry groups may lobby against the bill if they believe the compliance costs, despite the financial bars, still favor large incumbents. Their advocacy could lead to changes in the transparency obligations or the definition of 'covered developer' to reduce the burden on smaller firms.