Direct comparison
FRONTIER Act vs Voluntary AI Review
H.R. 9925 would mandate external AI audits every 6 months with unredacted access. Anthropic and OpenAI's current review practice is voluntary. Compared.
Written and maintained by CASRAI Editorial Board
Last updated
Ask CASRAI · free to try
Ask about FRONTIER Act vs Voluntary AI Review
Ask your first 2 questions free below. Subscribers get 150 a day for $29 a month.
Ask CASRAI answers research-administration questions and cites the passages behind every claim. When our sources don't cover a question, it says so.
Answers draw on CASRAI's guides and dictionary plus the federal and funder documents we index: Federal Register, Grants.gov, Regulations.gov and UKRI.
Works on this site and inside Claude, Cursor and the AI tools you already use.
Everything CASRAI publishes — this page, the dictionary, the guides and the news — stays free to read, with no account and no card.
How do FRONTIER Act (H.R. 9925, as introduced), Anthropic (Responsible Scaling Policy), OpenAI (Preparedness Framework) compare side by side?
The table below compares FRONTIER Act (H.R. 9925, as introduced), Anthropic (Responsible Scaling Policy), OpenAI (Preparedness Framework) across 6 procurement-relevant dimensions, from legal status through consequence for non-compliance.
Side-by-side comparison
| Dimension | FRONTIER Act (H.R. 9925, as introduced) | Anthropic (Responsible Scaling Policy) | OpenAI (Preparedness Framework) |
|---|---|---|---|
| Legal status | A bill introduced in the House July 23, 2026, referred to committee. Not enacted; not current law. | A voluntary company policy. No statute or regulator requires it or enforces its contents. | A voluntary company policy. No statute or regulator requires it or enforces its contents. |
| Who is covered | "Very large frontier developers": >$5B gross revenue and >$10B AI-development spend over the trailing 36 months (Sec. 2(22)), building "frontier models" trained above 10^26 operations (Sec. 2(12)). | Anthropic’s own models, whenever its RSP’s capability thresholds are triggered — scope set unilaterally by Anthropic’s own policy. | OpenAI’s own models, at OpenAI’s discretion — the framework’s own language is conditional ("when available and feasible"). |
| Who reviews | A licensed Independent Verification Organization (IVO), licensed under regulations issued per Sec. 3(c) — independent of both the developer and (per Sec. 3(c)(1)) required to disclose its own funding and self-audit for AI-industry ties. | External reviewers selected subject to approval by Anthropic’s own Long-Term Benefit Trust (LTBT) — a body Anthropic created and can amend via its own RSP revisions. | Unnamed "third parties," selected by OpenAI on a case-by-case basis; no standing external approving body is named in the public framework language. |
| Review cadence | Not less than once every 6 months (Sec. 5(f)(1)) — a statutory floor. | Tied to Risk Report cycles and capability-threshold triggers; no fixed calendar cadence stated in the RSP language quoted here. | No stated cadence; occurs "when available and feasible." |
| Materials access | "Unredacted materials, records, personnel, systems, and all other information reasonably necessary" (Sec. 5(e)(1)) — defined in statute. | Reviewers may see "unredacted sections" of the Risk Report, with every part covered by at least one reviewer — defined in Anthropic’s own policy, which Anthropic can revise. | Access terms not specified in the public framework language available to CASRAI. |
| Consequence for non-compliance | Civil penalty up to $1,000,000 per violation, each day a separate violation (Sec. 5(p)(2)(A)), enforced by the Attorney General and opt-in state AGs. | None external. A narrower RSP, or a skipped review, carries reputational cost but no fine, injunction, or regulator. | None external. Same as Anthropic — reputational only. |
Common questions
Common questions about FRONTIER Act (H.R. 9925, as introduced) vs Anthropic (Responsible Scaling Policy) vs OpenAI (Preparedness Framework)
Is the FRONTIER Act law right now?
+
No. H.R. 9925 was introduced in the House on July 23, 2026 and referred to committee. It has not passed either chamber and is not in force. Everything about its 6-month cadence, unredacted-access requirement, and penalties describes what the bill would require if enacted as introduced, not a current legal obligation on any lab.
What does the FRONTIER Act actually say about review frequency and access?
+
Quoting the bill text as introduced: Section 5(f)(1) requires the independent verification organization to "prepare and submit an assessment report not less frequently than once every 6 months." Section 5(e)(1) requires the developer to "grant an IVO timely access upon request to unredacted materials, records, personnel, systems, and all other information reasonably necessary" to the assessment. Both are drafted as fixed statutory floors, not negotiable terms.
Does Anthropic already do something close to what the FRONTIER Act would require?
+
Partially, and in Anthropic’s own words. Per anthropic.com/rsp-updates, RSP version 3.2 "authorizes the [Long-Term Benefit Trust] to request external review of Risk Reports, gives the LTBT the power to approve our selection of external reviewers," and version 3.4 clarifies that review "can involve multiple external reviewers reviewing different unredacted sections of the Risk Report, so long as all parts of the unredacted report are evaluated by at least one external reviewer." That is close to the FRONTIER Act’s access requirement. But the reviewer is approved by a body Anthropic itself created and can amend, not a federally licensed IVO, and there is no fixed 6-month statutory cadence in the policy language.
Would the FRONTIER Act, if passed, actually apply to Anthropic or OpenAI?
+
That depends on whether each company’s trailing-36-month revenue and AI-development spend cross the bill’s $5B/$10B thresholds (Sec. 2(22)) — figures CASRAI has not independently verified against a primary source for either company this session, so this piece does not assert either one currently qualifies. The thresholds are large enough that they are plausibly within reach of both companies’ scale, but "plausible" is not the same as confirmed.
What does NIKOLAI’s "shadow mapping" label mean on this page?
+
It means every row in NIKOLAI’s External Review crosswalk — including the FRONTIER Act row — is CASRAI’s own reading of a published document, not something the named organization, lab, or Congress has declared, endorsed, or been consulted on. NIKOLAI is CASRAI’s own independent, unendorsed reference dictionary, not a standard any of these organizations has adopted, and the FRONTIER Act’s binding requirement is a real legislative proposal that exists independently of, and is not endorsed by, NIKOLAI’s reading of it.
Where does NIKOLAI define "external review" itself?
+
As element N8.5 in NIKOLAI’s Transparency and Review track: "a record of an assessment performed by a party outside the model developer… capturing the review’s type, scope, and output." That element’s crosswalk is the richest one checked in this drafting pass — 11 shadow-mapped rows across 10 organizations and legal instruments, including Anthropic, OpenAI, the EU, California SB 53, METR, and the FRONTIER Act itself — and it is the definition this comparison uses to line up the FRONTIER Act, Anthropic, and OpenAI against the same three questions: type, scope, and output.
Going deeper







