Anthropic Proposes Federal AI Safety Bill Amid Controversy Over Compute Threshold

Anthropic has presented a draft federal AI safety bill to U.S. lawmakers, igniting discussions primarily centered around a specific compute threshold of 10²⁵ FLOPs. This figure is crucial as it determines which AI models are classified as 'frontier' under the proposed legislation.

Critics contend that the 10²⁵ FLOPs threshold reflects an outdated understanding of AI capabilities, potentially encompassing most enterprise-grade models. The bill builds on Anthropic's Advanced AI Framework, introduced in June 2026, and outlines that companies must exceed the 10²⁵ FLOPs benchmark and meet additional financial criteria to qualify as 'covered developers.'

• Companies must generate over $500 million in annual AI-derived revenue.
• They must also spend more than $1 billion annually on AI research and development.

Covered developers would be subject to various transparency obligations, including implementing formal safety frameworks and facilitating independent evaluations. However, the bill also limits government intervention, allowing the Commerce Department to restrict model deployment only under foreseeable catastrophic risks, defined as incidents causing damages exceeding $1 billion or resulting in more than 50 fatalities.

The debate over the 10²⁵ FLOPs threshold highlights a significant concern among critics regarding the potential advantages for larger AI labs, which can more easily absorb compliance costs compared to smaller firms. Furthermore, the static nature of the compute-based rules raises questions about their longevity as AI technology evolves.

As the bill progresses, observers are closely monitoring its potential impact on the competitive landscape for AI developers, particularly Anthropic, amid increasing regulatory scrutiny.