California SB 53
The first frontier AI law in force in the U.S. Publish a framework, publish transparency reports, report incidents, protect whistleblowers.
What it does
California's law, in force since January 2026. Big labs publish a safety framework, publish a report with each new model, report serious incidents within 15 days, and protect whistleblowers.
Applies to: Frontier models above 10^26 FLOP; most duties fall on developers with revenue above $500M.
- Large frontier developers publish and follow a frontier AI framework; update yearly; publish material changes within 30 days.
- Publish a transparency report before or at deployment of a new or substantially modified frontier model.
- Report critical safety incidents within 15 days; within 24 hours if imminent risk of death or serious injury.
- Send the state a summary of internal-use catastrophic-risk assessments every three months.
- No materially false statements about framework compliance. Whistleblower protections and an anonymous channel.
- Penalties up to $1M per violation, Attorney General only.
What's in it: Safety plan, Transparency report, Incident reports, Whistleblower protection
Where things stand
Anthropic, OpenAI, Google and xAI already published frameworks and system cards before the law took effect. OpenAI's Frontier Governance Framework explicitly maps its existing practices to SB 53. Nine months in, no lab has publicly attributed a delayed release to it.
Why it costs almost no lead
Its parts are paperwork labs already do, so the effect on the lead is under a day.
Biggest unknown: Whether the broad definition of deployment ever forces a transparency report before a trusted-access pilot.
Why it lowers p(doom) by ~0.17%
California's law: public safety plans, incident reports and whistleblower protections.
The strongest case that it costs more
The law's teeth are the duty to comply with your own framework and the ban on false statements about it. That converts a voluntary framework into a binding one, and the cost of a binding framework is whatever gates the lab wrote into it. If Anthropic's RSP thresholds had remained hard gates, SB 53 would have made them enforceable. Labs responded by softening their frameworks, which is a safety cost rather than a lead cost, but it shows the law has bite that this price does not capture.
The debate
For
- Anthropic endorsed the bill, 2025.
- Sen. Scott Wiener wrote the bill, which Gov. Newsom signed in September 2025.
Against
- Chamber of Progress said it still gets AI regulation wrong, 2025.
- OpenAI's Chris Lehane, CTA and a16z lobbied against it or raised objections, 2025.
Sources
Rough starting points, not precise forecasts. Lead costs assume China doesn't depend on U.S. models, the case least favorable to safety laws, and count 3 years. On the menu you can change every assumption and put in your own numbers. Last priced 2026-09-26.