Anthropic’s AI safety bill draws fire over its frontier compute line

Critics say the 10²⁵ FLOP threshold in Anthropic's federal proposal is outdated and could sweep in far more models than intended

Anthropic has handed US lawmakers a draft federal AI safety bill. The first fight is over a single number: 10²⁵ FLOPs.

That figure decides which AI models count as “frontier” under the proposal. Critics argue it reflects an earlier era of AI, and that it now captures most enterprise-grade models.

The proposal builds on Anthropic’s Advanced AI Framework, which the company introduced in June 2026. The draft legislation itself dates to October 2026.

Its central concept is the “covered developer.” To qualify, a company must train a model exceeding 10²⁵ FLOPs. It must also clear at least one financial bar.

The first bar is more than $500 million in annual AI-derived revenue. The second is over $1 billion in annual AI research and development spending.

Companies that meet the definition would face a set of transparency obligations. These include:

The bill also sets limits on government power. The Commerce Department could restrict a model’s deployment, but only when catastrophic risks are foreseeable. The draft defines those as incidents causing damages exceeding $1 billion or the loss of more than 50 lives.

Daily. Free. Join 34,000+ readers across crypto, finance, and policy.

Other agencies would be barred from halting deployment altogether.

By June 2025, at least 33 models from 12 developers had already passed the 10²⁵ FLOP mark. That count is projected to keep rising.

Comparisons with other rulebooks sharpen the point. California’s SB 53 sets its trigger at 10²⁶ FLOPs, ten times higher than Anthropic’s figure. Each step in exponent notation is a tenfold jump, so the gap is substantial.

The EU AI Act sits closer to Anthropic’s proposal. It uses 10²⁵ FLOPs as a rebuttable presumption, meaning a model at that level is presumed to carry systemic risk unless the developer can show otherwise. Anthropic’s version functions more like a hard compliance trigger.

There is a second line of criticism, and it pulls in the opposite direction. Some critics argue the draft is narrower and potentially weaker than competing proposals, such as the FRONTIER Act (H.R. 9925), particularly when it comes to rulemaking authority.

The incumbent-advantage argument is the most consequential critique on the table. Compliance work like safety frameworks, system cards and independent evaluations costs money and staff time. Large labs already maintain teams for much of this. Smaller players generally do not.

That said, the bill’s two-part test matters here. Crossing 10²⁵ FLOPs alone would not make a company a covered developer. It would also need more than $500 million in AI revenue or over $1 billion in AI R&D spending.

For lawmakers, the threshold debate exposes a structural problem with compute-based rules. Any fixed FLOP number ages quickly as hardware improves and training runs grow. A static line drawn today may cover dozens of models tomorrow, which is roughly what the June 2025 count already suggests.

The narrow Commerce Department mandate deserves attention too. Limiting intervention to catastrophic scenarios, with $1 billion in damages or more than 50 deaths as the markers, gives developers predictability. It also leaves regulators with little room to act on harms that fall short of catastrophe. That trade-off sits at the heart of the comparison with the FRONTIER Act.