Anthropic pitches Congress its own frontier-AI bill with a 50-death or $1B damage halt threshold

Anthropic has taken a frontier-AI bill of its own drafting to the United States Congress, according to a Dealroom report, in an effort to write its view of model risk into federal law. The proposal's most striking feature is a threshold for intervention: a model could only be halted once it had caused 50 deaths or $1 billion in damage.

The wording matters as much as the numbers. A harm threshold of that kind would tie government stop powers to consequences that can be counted and attributed, rather than to forecasts about what a model might do. For frontier labs, that is a lighter regime than pre-deployment licensing, but a more predictable one than no rules at all.

Proponents of threshold-based rules argue that clear triggers give regulators a defensible basis for action and give companies a line they can plan around. Critics are likely to ask whether 50 deaths or $1 billion in damage is set too high, since by the time losses reach that scale a halt looks like accountability after the fact rather than prevention.

The move also signals a shift in posture. Instead of waiting for legislators to define frontier-model risk, Anthropic is supplying drafting language itself, from risk classification to the conditions under which a model must be stopped. Whoever defines those categories early can shape how future compliance maps onto existing safety frameworks.

That is also where the tension lies. A model developer is both an author of the rules and an object of them, and lawmakers will have to decide whether a threshold written with industry input protects the public or preserves room to operate. The reported figures will be read as a negotiating position as much as a safety proposal.

Three things are worth watching: whether congressional committees take up the text, whether the 50-death and $1 billion figures survive or are revised in review, and whether rival labs publish competing drafts. Until a threshold is enacted, the proposal is better understood as an opening bid than an endpoint.