Eight proposals built on one design, and how they differ on who checks, who can stop a model, and what states keep
The most common design for frontier-model regulation began as a California transparency law and now runs through two state laws that adopted it, a House discussion draft, an introduced House bill, an unpublished Senate draft, and two industry proposals. Every member shares the same skeleton: covered developers write and follow a safety framework, publish a report when they release a new model, and report critical safety incidents on a short clock. Catastrophic risk is typically defined by a casualty-or-damage threshold: more than 50 deaths or $1 billion from a single incident in SB 53, the FRONTIER Act, and Anthropic's text.
What changes from one version to the next is what is built on top of that skeleton. Read in order, the table shows three additions: someone outside the developer checking its claims, someone able to stop a model in an emergency, and, in the federal versions, preemption of exactly the state laws the design began as.
In order of when each entered the debate. Scroll sideways on a small screen.
| Proposal | Form | Coverage | Who checks | Incident clock | Who can stop a model | Preemption |
|---|---|---|---|---|---|---|
| CA SB 53Sep 29, 2025 | State law, in force | Large frontier developers above $500M in revenue | No one — developers describe their own risk management | 15 days; 24 hours if imminent harm | No one | State law |
| NY RAISEDec 19, 2025 | State law, in force | Frontier developers above $500M in revenue | No one — published safety and security protocols | 72 hours | No one | State law |
| OpenAI BlueprintJun 2, 2026 | Industry blueprint (no text) | Large frontier developers (thresholds not specified) | Annual third-party audits, plus a mandatory CAISI evaluation before release that is advisory only | Required; clock not specified | No one — CAISI evaluates and recommends but cannot block, and a developer may ship if CAISI misses its deadline | Seeks it: codify SB 53, RAISE, and SB 315 federally, then preempt state frontier-safety laws; states keep youth protection and energy |
| GAAIAJun 4, 2026 | House discussion draft; its frontier title became the FRONTIER Act | Large frontier developers above $500M in revenue | Independent Verification Organizations, licensed by CAISI, audit developer frameworks | 15 days; 24 hours if imminent | No one | Narrow: state laws targeting model development only, sunsetting after three years |
| Illinois SB 315Jul 6, 2026 | State law, effective Jan 1, 2027 | Above 10^26 operations and $500M in revenue | Annual independent compliance audit from 2028; quarterly internal-use risk summaries to the state | 72 hours; 24 hours to law enforcement if lives are at risk | No one | State law; accepts a federal regime with equivalent incident reporting and audits as satisfying it |
| FRONTIER ActJul 23, 2026 | Introduced House bill (H.R. 9925) | All developers above 10^26 operations; heavier tiers at $50M revenue + $1B AI spending, and at $5B + $10B | Annual compliance audit; for the largest tier, a federally licensed verifier judging every six months whether mitigation is adequate | 72 hours; 24 hours to law enforcement if imminent | The Commerce Secretary, by emergency order on a finding of imminent catastrophic risk (45-day provisional, 90-day renewable final), reviewable in D.C. District Court | Permanent, limited to frontier risk transparency, third-party auditing and verification, and incident reporting |
| Thune–KlobucharSep 11, 2026 | Unpublished Senate draft; stalled | Frontier developers (thresholds not reported) | Developers test their own models and present results to Commerce | Prompt notice of significant capability advances (incident clock not reported) | A federal court, on Commerce's application, for a developer that fails its risk-management obligations | State catastrophic-risk laws; scope contested |
| Anthropic Draft BillOct 16, 2026 | Bill text drafted by Anthropic; no sponsor | All developers above 10^25 operations; large developers above $500M in AI revenue or $1B in AI spending | A registered independent evaluator every 180 days, with published reports; findings are advisory | 15 days; 24 hours if imminent | The D.C. District Court, on application by the head of Commerce, for 45 days at a time, if Commerce proves imminent risk through one of four named pathways and that nothing less restrictive would work | Disclosure, third-party evaluation, and incident reporting, for every AI developer; void if the federal requirements fall |
Same components — federal standards, independent audits, incident reporting — but the audit is a gate: no release until it confirms the standards are met. Opposes the preemption the federal members carry.
California building the verification layer onto SB 53 itself: certified verifiers embedded in labs, independent verification of SB 53 filings, and a verified kill switch.
Only the emergency brake, without the framework-and-report skeleton, and placed on whoever operates the model rather than whoever trained it.
A different route to verification: instead of third-party evaluators, a federal board gets the model and its weights 45 days before release and sets binding standards — but cannot block a release.