Anthropic's policy ask for the AI exponential

Anthropic published a policy proposal arguing that the current policymaking process moves too slowly for AI's pace, and laying out what it thinks regulation of frontier systems should look like. The piece sits in two parts. The first, an Advanced AI Framework, would give the government authority to block or deter dangerous deployments of frontier models, with civil penalties tied to a developer's global annual revenue. The duties only kick in above a threshold: models trained with more than 10^25 FLOPs, from companies with at least $500 million in AI revenue or $1 billion in AI R&D spending. The second part addresses the economic side, with proposals for how the workforce and public finances should adapt as capable AI spreads through the economy.

Anthropic names four categories of catastrophic risk the framework is meant to handle: biological weapons, cyber attacks on critical infrastructure, loss of control over the systems themselves, and automated R&D in which AI accelerates its own development without human checks. Concrete obligations for covered developers include publishing safety testing summaries, engaging independent evaluators to review those findings, securing model weights and training infrastructure to a high standard, and keeping safety frameworks and system cards open to public review. The framing is unusual for an AI lab: Anthropic is asking to be regulated, with thresholds high enough to land mostly on itself and a handful of peers, and it cites its race-to-the-top thesis as the reason.

Why it matters

If you build on frontier APIs, this is the company most likely to shape the rules you and your customers will eventually live under, so it pays to know exactly what Anthropic is asking for. The 10^25 FLOPs and revenue floors are the most concrete claim here, and they tell you where the line between "covered" and "uncovered" model providers may end up being drawn.