None of the six federal and state measures we read requires OpenAI, Anthropic or any other AI developer to get government approval before releasing a model: 0 of 6, by our count. The closest thing on the books is voluntary. Executive Order 14409, signed on 2 June 2026, told agencies to design a voluntary framework under which developers could give the government access to “covered frontier models” for up to 30 days before release to other trusted partners.
Section 3 of the same order says nothing in it shall be construed to authorize “a mandatory governmental licensing, preclearance, or permitting requirement” for new AI models. The question comes from a post on X by roon (@tszzl) at 00:37 UTC on 16 November, which showed 85,080 views when we read it at about 06:05 UTC. His view, verbatim: “people have little understanding of how much the federal government is already involved in the ai industry. for example if there’s a new model you need to convince the feds that new model is safe before launching it or men with proverbial guns and export control bans show up”.
Business Pill · WHEN TO TELL
A one-minute explainer of the notification clock: the days between finding a problem and telling the people it affects. It teaches the general idea only and says nothing about any company, law or order in this story.
The key insight: As we read it, the federal role before a U.S. model launch is an offer, not a gate: 0 of the 6 measures we read requires approval, by our count, while Executive Order 14409 sets out a voluntary window of up to 30 days, and the model evaluations NIST has published since January 2025 cover PRC-developed models only, 6 of 6.
What the Federal Rules Ask Before a Launch
Section 3 of Executive Order 14409, titled “Secure Frontier Model Deployment”, gives the Treasury, the Department of War through the National Security Agency, and Homeland Security through CISA 60 days to act, in consultation with others including the Commerce Department through NIST.
They are to build a classified benchmarking process for the advanced cyber capabilities of AI models and set the threshold at which a model becomes a “covered frontier model”, a call the order gives to the Director of NSA. They are also to design the voluntary framework: developers “would be able to” ask whether a model in development meets that bar, give the government access for up to 30 days before release to other trusted partners, and help pick the trusted partners that get early access.
America’s AI Action Plan of July 2025 puts the same idea as a recommended action: “Evaluate frontier AI systems for national security risks in partnership with frontier AI developers, led by CAISI at DOC in collaboration with other agencies with relevant expertise in CBRNE and cyber risks.” It is a plan, published by the White House, not a rule binding on developers.
Executive Order 14365 of 11 December 2025 deals with the states: it told the Attorney General to set up an AI Litigation Task Force to challenge State AI laws inconsistent with its policy, and asked the FCC to consider “a Federal reporting and disclosure standard for AI models that preempts conflicting State laws”.

What Was Posted on X
roon (@tszzl) posted this on 16 November at 00:37 UTC; it showed 85,080 views when we read it about five and a half hours later. It argues that a new model needs the federal government’s safety sign-off before launch. This piece checks that view against the executive orders, NIST’s testing record, export rules and the California and New York laws, and finds the federal track on paper is voluntary.
The Voluntary Testing Track at NIST
The testing relationship predates these orders. On 29 August 2024 the U.S. AI Safety Institute at NIST announced agreements with Anthropic and OpenAI, under which each company’s Memorandum of Understanding lets the institute “receive access to major new models from each company prior to and following their public release”. NIST’s page notes that in June 2025 the institute was re-established as CAISI by Commerce Secretary Howard Lutnick.
NIST’s own news list for the center, 37 items long, shows two published pre-deployment evaluations of U.S. lab models, both joint with the UK institute: Anthropic’s upgraded Claude 3.5 Sonnet (19 November 2024) and OpenAI’s o1 (18 December 2024).
Since then, the list shows six published evaluations or assessments, all of models from developers NIST describes as PRC-based or from the People’s Republic of China: DeepSeek (30 September 2025), Kimi K2 Thinking (12 December 2025), DeepSeek V4 Pro (1 May 2026), GLM-5.2 (17 July 2026), Kimi K3 with the UK institute (23 July 2026) and GLM-5.3 (17 September 2026). That is 6 of the 6 model evaluations it lists since January 2025, by our count.
A 25 September 2025 note on the same list says CAISI worked with OpenAI and Anthropic “to identify security issues with their advanced AI systems”.
The labs describe that work in their own posts. Anthropic says that its “voluntary work together began as initial consultations” and that CAISI and the UK institute tested its Constitutional Classifiers “on models like Claude Opus 4 and 4.1 prior to deployment”. OpenAI says it was “among the first companies to enter into voluntary agreements” with both bodies, and that CAISI found two vulnerabilities in ChatGPT Agent, “fixed by OpenAI within one business day”.
The center’s NIST page now carries the name Center for Advancing Innovation and Standards for Super Intelligence (CAISSI), the wording the Super Intelligence executive order asks agencies to use, and lists as a task: “Establish voluntary agreements with private sector SI developers and evaluators”.

Export Controls: Chips, Not Launch Permits
On 13 May 2025 the Commerce Department announced a rescission of the AI Diffusion Rule, issued on 15 January 2025 with compliance requirements that were set to take effect on 15 May 2025. BIS said it “plans to publish a regulation formalizing the rescission and will issue a replacement rule in the future”; the Federal Register search we ran on 16 November returned no BIS document formalizing it.
The same release added export licensing requirements for chips shipped to Malaysia and Thailand, citing known diversion. The export rules in force we read apply to chips; none of the documents we read makes an export licence a condition of releasing a model. Fifteen chip export bills sit in Congress, none of them law, by our earlier count.
What California and New York Require
California’s SB 53, signed into law on 29 September 2025, covers models trained with more than 10^26 operations. “Before, or concurrently with” deploying a new frontier model, a developer must publish a transparency report; a large frontier developer, one that with its affiliates had more than $500 million in annual gross revenue the year before, must add summaries of its catastrophic-risk assessments and their results.
Critical safety incidents go to the state’s Office of Emergency Services within 15 days. Penalties reach $1,000,000 per violation, in a civil action brought only by the Attorney General. The provisions we read ask for publication and reporting, not approval.
New York’s RAISE Act was signed on 19 December 2025 as chapter 699. Per Governor Kathy Hochul’s office, large developers must publish information about their safety protocols and report incidents within 72 hours, with an oversight office in the Department of Financial Services and penalties of up to $1 million for a first violation and $3 million for later ones.
Pre-release access is being proposed in Congress, not required: a Senate bill from Mark Warner would give a new federal board access to frontier models, weights included, 45 days before release, and Senator Maria Cantwell’s framework says covered models should not ship before an independent audit.
The Structural Read
As we read roon’s post, the step he describes, convincing the government before a launch, matches the voluntary track on paper: the 2024 agreements give the government access to major new models before public release, and Executive Order 14409 asks agencies to design a framework for that access while ruling out mandatory preclearance.
As we read the two state laws, California and New York regulate what the largest developers must disclose, not whether they may ship: SB 53’s transparency report is due “before, or concurrently with” deployment, and the RAISE Act adds a 72-hour incident clock, per Governor Hochul’s office.
Executive Order 14409, Sec. 3(c) (2 June 2026)
“Nothing in this section shall be construed to authorize the creation of a mandatory governmental licensing, preclearance, or permitting requirement for the development, publication, release, or distribution of new AI models, including frontier models.”
Three Implications
FOR AI DEVELOPERS Under the measures we read, no federal licence is needed to release a model; what binds are state disclosure and incident rules for the largest developers, with penalties of up to $1,000,000 per violation in California.
FOR THE POLICY TRACK Executive Order 14365 tells an AI Litigation Task Force to challenge State AI laws that conflict with its policy, while a Senate bill from Mark Warner would add pre-release access 45 days before launch.
RELATED Our Super Intelligence order piece explains why NIST’s center now carries a new name.
The Business Engineer Lens
This story maps onto the Business Engineer framework Inside Anthropic’s Permission Layer.
The framework says: “The governance layer sits at the intersection of capability, regulation, national security, access control, and geopolitical alignment.”
As we read it, the U.S. measures read for this piece place that governance layer mostly in voluntary testing agreements and state disclosure laws, not in a federal launch gate. The framework does not discuss Executive Order 14409 or roon’s post.
What Is Not Established
Whether the agencies have finished the benchmarking process and the voluntary framework Executive Order 14409 asked for within 60 days is not stated in the documents we read, nor whether any developer has signed up to it.
Which 2026 model releases, if any, went through CAISI testing before launch is not shown in NIST’s news list. Roon’s post names no rule, and we did not verify his employer. We did not contact roon, OpenAI, Anthropic, NIST, the White House or BIS for this piece.
The Bottom Line
On the record, a U.S. model launch needs no federal sign-off: 0 of the 6 measures we read requires one, by our count. What exists is a voluntary track, Executive Order 14409’s up-to-30-day access framework and CAISI’s testing agreements, alongside publish-and-report laws in California and New York and export rules aimed at chips.
How the administration is using industrial policy instead is in our national champions piece.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. This explainer rests on primary texts: Executive Orders 14409, 14365 and 14434 as printed in the Federal Register; America’s AI Action Plan PDF from whitehouse.gov; NIST’s CAISI pages and its 37-item news list for the center, both pages read on 16 November 2026; the Bureau of Industry and Security’s 13 May 2025 release and a Federal Register search we ran on 16 November; California’s chaptered SB 53; the New York Senate’s bill page for S6953B and Governor Hochul’s release; and Anthropic’s and OpenAI’s own posts on their CAISI work.
Roon’s post was read in full through a public X mirror. The six-measure tally and the evaluation counts are our count from those documents. Earlier bills and orders are linked to our own reporting. We did not contact roon, OpenAI, Anthropic, NIST, the White House or BIS for this piece.
Sources: Federal Register, Executive Order 14409 · Federal Register, Executive Order 14365 · The White House, America’s AI Action Plan · NIST, CAISI news list · NIST, CAISSI page · Bureau of Industry and Security, 13 May 2025 · California SB 53 (Chapter 138) · New York Senate, S6953B · Anthropic · OpenAI · roon on X









