Anthropic AI Safety Bill Faces Criticism Over 10²⁵-FLOP Frontier Threshold
Key Takeaways
- •Anthropic's draft would classify a company as a covered developer only if it trains a model above 10²⁵ FLOPs and also surpasses $500 million in annual AI revenue or $1 billion in annual AI research and development spending.
- •Covered developers would be obligated to implement formal safety frameworks, publish system cards, enable independent evaluations, and report safety incidents.
- •Under the proposal, the Commerce Department could intervene against a model's deployment only when catastrophic risks are foreseeable, with catastrophe defined as more than $1 billion in damage or more than 50 lives lost.
- •By June 2025, at least 33 models from 12 developers had already crossed the 10²⁵-FLOP mark, feeding criticism that the threshold reflects an earlier stage of AI development and could capture most enterprise-grade models.
- •The draft's trigger sits ten times below California's SB 53 (10²⁶ FLOPs) and diverges from the EU AI Act's presumption-based approach, while some critics consider it narrower and potentially weaker than the FRONTIER Act on rulemaking authority.
Anthropic has given US lawmakers a draft federal AI safety bill, with the first major dispute centered on a single number: 10²⁵ FLOPs.
That threshold would determine which AI models qualify as “frontier” systems under the proposal. FLOPs, or floating-point operations, measure the computing power used to train a model — and wherever that line is set decides which systems, and which developers, fall inside a compliance perimeter and which stay outside it. Critics say the figure reflects an earlier stage of AI development and could now encompass most enterprise-grade models.
What the draft proposes
The proposal builds on Anthropic’s Advanced AI Framework, introduced by the company in June 2026. The draft legislation itself is dated October 2026.
Its central concept is the “covered developer.” A company would qualify if it trained a model exceeding 10²⁵ FLOPs and met at least one of two financial thresholds.
The first threshold is more than $500 million in annual AI-derived revenue. The second is more than $1 billion in annual AI research and development spending.
Companies meeting the definition would be subject to several transparency requirements, including:
- Implementing formal safety frameworks
- Publishing system cards documenting how models work and how they were tested
- Facilitating independent evaluations
- Reporting safety incidents
The bill would also limit the government’s authority. The Commerce Department could restrict a model’s deployment only when catastrophic risks are foreseeable. The draft defines catastrophic incidents as those causing more than $1 billion in damage or resulting in the loss of more than 50 lives.
Other agencies would be prohibited from halting deployment altogether.
Why the 10²⁵-FLOP threshold is contested
By June 2025, at least 33 models from 12 developers had already exceeded the 10²⁵-FLOP mark. That number is expected to continue increasing.
Comparisons with other regulatory frameworks underscore the debate. California’s SB 53 sets its trigger at 10²⁶ FLOPs, which is ten times higher than Anthropic’s threshold. Each increase of one in exponent notation represents a tenfold change, making the difference significant.
The EU AI Act is closer to Anthropic’s approach. It uses 10²⁵ FLOPs as a rebuttable presumption: a model reaching that level is presumed to pose systemic risk unless its developer demonstrates otherwise. Anthropic’s proposal would operate more like a mandatory compliance trigger.
Criticism also comes from the opposite direction. Some critics say the draft is narrower and could be weaker than competing proposals, including the FRONTIER Act (H.R. 9925), particularly regarding rulemaking authority.
Implications for startups, established labs and lawmakers
The potential advantage for incumbent companies is one of the most consequential criticisms. Compliance measures such as safety frameworks, system cards and independent evaluations require money and staff time. Large AI labs already maintain teams responsible for much of this work, while smaller companies generally do not.
However, the bill’s two-part test is significant. Exceeding 10²⁵ FLOPs alone would not make a company a covered developer. The company would also need more than $500 million in AI revenue or more than $1 billion in AI research and development spending.
For lawmakers, the debate highlights a structural problem with compute-based rules. Any fixed FLOP threshold can become outdated as hardware improves and training runs expand. A line established today could encompass dozens of models tomorrow, which is broadly consistent with the number of models that had already crossed the threshold by June 2025.
The Commerce Department’s limited mandate is also significant. Restricting intervention to catastrophic scenarios, defined by more than $1 billion in damage or more than 50 deaths, gives developers greater predictability. At the same time, it leaves regulators with little scope to address harms that fall short of catastrophe. That trade-off is central to the comparison with the FRONTIER Act.
The trigger level, the two-part financial test and the scope of Commerce’s rulemaking authority are therefore the points where Anthropic’s draft diverges from the alternatives already on the table — California’s higher 10²⁶-FLOP trigger, the EU AI Act’s presumption-based approach and the FRONTIER Act that some critics consider stronger on rulemaking.
Source: CryptoBriefing
