RAISE Act vs SB 53: how New York and California frontier AI laws differ

California's SB 53 and New York's RAISE Act are the two US state laws aimed squarely at frontier AI developers, and they are close cousins with different teeth. Both require published safety frameworks and incident reporting for developers training models above 10^26 operations. New York's amended act reports faster, can fine harder on repeat violations and adds a financial regulator. California's has whistleblower protections and has already been in force since January 1, 2026.

The short version

Here is the quick comparison before the detail. California's Transparency in Frontier AI Act, SB 53, was signed in late September 2025 and became operative January 1, 2026, according to FPF's January 8, 2026 analysis. New York's RAISE Act, S6953B, passed the state Senate 58 to 1 on June 12, 2025 and was signed as Chapter 699 on December 19, 2025, per the New York Senate's bill page. A chapter amendment then aligned it with California. Cooley's March 31, 2026 analysis puts the amended act's effective date at January 1, 2027.

Who is covered

Both laws start from frontier models, defined by training compute above 10^26 operations. Cooley says New York's amendment dropped the original compute-cost tests and now counts fine-tuning and reinforcement learning toward the threshold. A "large frontier developer" is one whose affiliates together had annual gross revenue above $500 million in the prior calendar year.

That revenue test replaced a different one in New York. The original bill defined large developers as those spending over $100 million on compute for frontier model training in aggregate, according to the Senate page. The change matters for well-funded startups that spend heavily on compute but earn little.

What developers must do

Both laws rely on transparency. Developers publish a frontier AI framework describing how they assess and mitigate catastrophic risk, and publish transparency reports before deploying new frontier models. FPF notes the catastrophic risk standard is shared: 50 or more deaths or serious injuries, or $1 billion or more in damages, plus chemical, biological, radiological or nuclear uplift, criminal conduct or cyberattacks without meaningful human intervention, and evading developer or user control.

Both drop what earlier drafts carried. Cooley says each law dropped the independent audit mandate and the deployment prohibition. New York's original bill barred deploying a model posing an unreasonable risk of critical harm, and the chapter amendment removed it. The final structure is disclosure, not permission.

Where they differ

Incident reporting is the sharpest gap. New York requires a report within 72 hours of forming a reasonable belief that an incident occurred, and 24 hours if there is imminent danger of death or serious injury. California's baseline is 15 days, with a 24-hour window for imminent risk.

Penalties differ in structure. Under California's law the cap is $1 million per violation. New York's amended act sets $1 million for a first violation and $3 million for later ones, down from $10 million and $30 million in the version the Senate passed.

Governance differs too. FPF's table says New York requires large developers to file a disclosure statement with the Department of Financial Services before operating a frontier model in the state, renewed every two years, with a fee. Cooley says oversight moved to a new office within DFS with broad rulemaking power, which California's law lacks. The Attorney General brings civil actions in New York.

Two items run the other way. California's law includes whistleblower protections, and New York's, per FPF and Cooley, does not include frontier-specific ones. And New York is limited to models developed, deployed or operating in the state, with exemptions for accredited academic institutions and the Empire AI consortium.

Private suits

The RAISE Act "expressly does not create a private right of action," FPF writes, quoting the act. Enforcement therefore rests with the Attorney General. FPF does not address private suits under California's law, so we do not state a position on it.

What else is moving in the states

The Transparency Coalition's October 16, 2026 update counts 85 new AI laws in 27 states this year. Frontier-specific bills are still in play. New Jersey's Assembly committee hears A 5275 on October 19, on developer disclosures to the Attorney General. Michigan's HB 4668, on safety standards for certain AI developers, passed the state Senate on April 29 and sits in a House committee. Massachusetts has proposed tough rules in the Senate's version of an economic development bill, WBUR reported on October 16, and that bill is still under negotiation.

Congress is circling the same ground. We track the federal AI safety bills separately, and federal preemption of state laws is a live risk to everything above.

What to do with this

For a developer above the threshold, the practical rule is to plan to the stricter law. That means a 72-hour incident process, a published framework and a transparency report at each release, plus a registration with New York's financial regulator before the January 2027 start. For everyone else, the useful thing is to read the model's published framework and ask whether the incident clock it describes matches the law that applies.

Our take

These laws are modest by design. They ask labs to say what they will do, report what goes wrong, and be fined if they lie. That is more than the federal government currently requires, and it is the baseline the next bills will be measured against. We would watch two details: the rules New York's financial regulator writes, and whether federal preemption arrives before 2027.