In This Article
Key Takeaways
- On July 23, 2026, Rep. Jay Obernolte (R-CA-23) introduced H.R. 9925, the FRONTIER Act, with five original cosponsors. It was referred to committee the same day.
- The text posted by the sponsor's office defines a “frontier model” by compute — more than 1026 integer or floating-point operations — and then tiers obligations by a developer's revenue and AI development spending.
- Every covered developer would publish a transparency report per model. Large developers add a published safety framework and an annual independent compliance audit; very large developers add a licensed verification organization.
- Critical safety incidents would be reported to a new Under Secretary of Commerce for AI Security within 72 hours; a separate 24-hour clock covers notifying law enforcement of imminent-harm incidents.
- Nothing has passed. GovTrack lists the bill's status as introduced, with no markup, vote or cost estimate recorded.
Congress spent 2026 circling frontier AI with frameworks and discussion drafts. On July 23, 2026, a piece of it became a bill. Rep. Jay Obernolte (R-CA-23) introduced H.R. 9925, the Frontier Risk Oversight, National Transparency, Independent Evaluation, and Reporting Act — the FRONTIER Act — with five original cosponsors: Reps. Lori Trahan (D-MA-3), Erin Houchin (R-IN-9), Scott Peters (D-CA-50), Scott Franklin (R-FL-18) and Suhas Subramanyam (D-VA-10), per GovTrack.
What was introduced, and when
The sponsors' announcement says the bill “was developed as part of the broader Great American AI Act framework.” It was referred the same day to Energy and Commerce and to Science, Space, and Technology. Everything below comes from the text posted by the sponsor's office: no official text is yet on congress.gov, and GovTrack still reads “Text: Not available yet.”
Who the bill would cover
Coverage starts with arithmetic rather than adjectives. Sec. 2(12) defines a frontier model as a foundation model trained using more than 1026 integer or floating-point operations — the original training run plus any later fine-tuning, reinforcement learning or substantial modification. Two size tiers build from there, measured with affiliates over 36 months.
Tiers as defined in the bill text
| Tier | Threshold (36-month window) | Core duty added |
|---|---|---|
| Frontier developer | Trained or began training a model above 1026 operations | Publish a transparency report per model |
| Large frontier developer | > $50 million gross revenues and ≥ $1 billion AI-related development spend | Published safety framework, registration, annual independent audit |
| Very large frontier developer | > $5 billion gross revenues and ≥ $10 billion AI-related development spend | Ongoing assessment by a licensed verification organization |
Rep. Franklin, in the sponsors' release, said the bill focuses “only on the handful of companies developing the most powerful frontier AI models, not the startups and innovators driving America's AI ecosystem.” Bloomberg Government's headline coverage named OpenAI and Anthropic as the sort of developers in scope.
The duties, tier by tier
Every frontier developer, whatever its size, would publish a transparency report before or alongside deploying a new or substantially modified model: release date, capabilities, intended uses, restrictions and risk assessments, with redactions allowed for trade secrets and security. The sponsors' release calls this a “model card”; that phrase appears nowhere in the text, which says “transparency report” — though model cards, datasheets and AI impact assessments are already the governance documentation federal AI programs are asked to produce under NIST AI RMF and OMB M-25-21.
Large developers carry more: publish a frontier AI framework within one year of enactment (or 90 days after first qualifying), justify any material modification within 30 days, and commission an independent compliance audit on that clock “and annually thereafter.” Auditor independence is statutory — no financial interest either way, and payment may not depend on the findings. Very large developers would also retain an Independent Verification Organization licensed by a new Under Secretary of Commerce for AI Security, whom the Secretary appoints.
Incident reports and emergency orders
Two clocks, often conflated, run on different tracks. Sec. 4(h)(1) gives a developer 72 hours to report a critical safety incident to the Under Secretary after learning facts sufficient to form a reasonable belief one occurred. Sec. 4(h)(2) gives 24 hours to notify law enforcement when an incident poses imminent risk of death or serious injury. The Washington Examiner reported the deadline as 24 hours after discovery; where an account and the text diverge, we follow the text.
Both terms are narrow. A “critical safety incident” means exfiltration of model weights, harm from a materialized catastrophic risk, loss of control, or a model deceiving its developer to subvert controls. “Catastrophic risk” means a foreseeable, material risk of contributing to more than 50 deaths or serious injuries, or more than $1 billion in property damage, from one incident.
Sec. 8 would let the Secretary of Commerce suspend or restrict development, deployment or internal use of a model on a written finding of imminent catastrophic risk. Those orders are time-boxed — provisional orders lapse after 45 days, final orders after 90, renewable only on a fresh finding — with an expedited hearing at which the Secretary bears the burden. Penalties run to $1 million per violation per day under Secs. 4 and 5, and $10 million per day for defying an emergency order.
The state-law clause
Sec. 9 would bar states from imposing new substantive obligations on AI developers in three defined “Covered Subject Areas”: frontier AI risk transparency, third-party auditing and independent verification, and incident reporting. A rule of construction preserves generally applicable laws, state regulation of how deployers and users use AI, laws protecting minors, and procurement rules. The sponsors' section-by-section says the clause “is aimed at closely related State frontier-safety statutes (e.g., CA SB-53, NY RAISE, IL SB-315)” — laws we have not reviewed.
One correction, because it is circulating: the Washington Examiner describes “a three-year ban on states implementing regulations on the development of AI models.” No three-year term, sunset or moratorium appears in the introduced text. That preemption was reported in June 2026 against the broader Great American AI Act draft — a different document.
Why it matters
The following is our analysis, not reported fact. For most practitioners the direct burden is zero: a $1 billion development-spend floor puts these duties on a very short list of companies. The second-order effects are what to watch.
A compute threshold is a moving target that vendors, not regulators, control — and the bill lets the Under Secretary raise the compute, revenue and spending numbers by rule. If something like this became law, the artifacts change: published safety frameworks, transparency reports and redacted audit reports become procurement diligence you can cite — and the work of producing them is versioning the risk register, the measurement plan and the test results alongside the model, not writing a document at the end. Licensing verification organizations would also create a professional audit market where none exists, and the bill orders an annual GAO report on whether those auditors stay independent. If your own evaluation practice is thin, close the gap between “we tested it” and “we can evidence it” — see our guides to testing agents and hallucination rates.
Then the caveat: bills die. This one has a bipartisan sponsor list, two committees of referral, no markup scheduled and no cost estimate; GovTrack's model put its odds of enactment at 4%. Rep. Subramanyam, quoted in the release, made the case anyway: it is “an important, bipartisan step in the right direction.”
A note on one quote we did not build on
Rep. Houchin's statement refers to an AI system that “broke out of its own developer's testing environment” earlier that week. That is her characterization; we have not verified which incident she meant.
Track what Congress passed, not what it proposed
Our federal AI coverage separates enacted rules from introduced bills and executive-branch frameworks.
Read the federal AI recapSources: H.R. 9925 — Congress.gov; Rep. Obernolte press release (July 23, 2026); FRONTIER Act bill text (PDF, sponsor's office); Section-by-section (PDF); Rep. Trahan press release; GovTrack — H.R. 9925; Washington Examiner; Bloomberg Government (headline only; body paywalled); Roll Call (June 4, 2026, on the earlier draft). Analysis and framing by Precision AI Academy.
Common questions
Is the FRONTIER Act law? No. It was introduced and referred to committee on July 23, 2026, and as of this writing there is no markup, floor action or published cost estimate.
Would it apply to my company? Almost certainly not, unless you train models above 1026 operations. The heavier duties also require more than $50 million in revenue and at least $1 billion in AI development spending.
Does it ban state AI laws? Not broadly. Sec. 9 preempts new state obligations on developers in three named areas, and preserves generally applicable law, use-and-deployment regulation, laws protecting minors, and procurement rules.
Who would enforce it? The U.S. Attorney General, and state attorneys general who opt in. Emergency-order authority sits with the Secretary of Commerce.