President Trump's June 2 executive order on artificial intelligence — formally titled "Promoting Advanced Artificial Intelligence Innovation and Security" — directed federal agencies to build a framework under which AI developers provide the government access to "covered frontier models" for up to thirty days before those models reach other customers [1]. The order frames the process as voluntary. It also expressly states it is not a licensing or preclearance regime. Both statements are technically accurate and together obscure what the framework actually does.

This paper's July 2 reporting on Anthropic's model-access channel running through Commerce established that government approval is now an operational variable in frontier-model deployment — Commerce licensed, then suspended, then restored Fable 5 and Mythos 5, making the agency a standing gatekeeper for what Anthropic ships to customers. The 30-day preview framework is the policy architecture that generalizes that variable across the entire frontier-AI industry.

The framework's load-bearing term is "covered frontier models." The executive order directs agencies to collaborate with AI developers to design the voluntary process, which means the definition of what constitutes a "covered" model — and therefore what falls within the 30-day window — is being written by the companies most likely to benefit from setting the threshold as high as possible [1]. OpenAI, Google, and Anthropic are not negotiating a safety standard. They are negotiating which of their smaller competitors must also submit to federal review.

An earlier draft of the order allowed the government 90 days of preview access. The final version reduced that to 30 days [2]. The reduction came after industry lobbying and reflects the product-cycle reality that a 90-day federal hold would effectively give competitors three additional months to close capability gaps. Thirty days is long enough to identify a security vulnerability; it is also short enough that a well-capitalized incumbent can plan around it while a resource-constrained startup cannot.

Five major frontier labs have already accepted federal review obligations under the Commerce Department's CAISI program [3]. Meta remains the lone holdout. The significance of that holdout is not philosophical: Meta's open-weight strategy — releasing model weights publicly — is structurally incompatible with a federal preview window, because a model whose weights are published cannot be recalled once it leaves the facility. The 30-day framework, applied to open-weight releases, would function as a de facto ban on public weight releases from covered models, which is precisely why the "voluntary" framing matters for Meta and no one else at the frontier.

The NSA and CISA have been directed to develop a classified benchmarking process for identifying AI models with advanced cyber capabilities — which is the actual security instrument the order creates [1]. A model cleared by that classified benchmark can, in principle, move through the 30-day window with government confidence behind it. A model that fails the classified benchmark faces a situation the order does not describe: the order says the review is not preclearance, but it does not say what happens when the classified process returns an adverse finding.

The Commerce Department's conduct in the Fable-Mythos case supplies the answer the order omits. When Amazon researchers flagged a jailbreak enabling vulnerability-exploitation code in June, Commerce did not invoke a formal export-control process first; it issued an emergency suspension that took effect in hours [4]. The 30-day preview window exists precisely to prevent that kind of abrupt action. But it also creates a formal channel through which a similar adverse finding could arrive more deliberately and with more institutional cover. Voluntary entry into a 30-day federal review is not the same thing as entering a 30-day federal review with an express legal right to exit it.

What the order creates is a new operative category: the model in preview. A covered frontier model in the 30-day window is no longer fully the company's to deploy but is not yet subject to formal regulatory hold. It exists in a legal condition the order does not name. That unnamed condition is already shaping release operations: reporting confirms that the framework is affecting deployment timelines at participating labs, with models being prepared for federal access in ways that did not exist six months ago [2].

The threshold negotiation will resolve one thing: who has the leverage to set the terms. It will leave unresolved the question the order does not answer — what happens when the government, having seen a model in preview, wants changes the developer declines to make. The order says this is not preclearance. The Fable-Mythos precedent demonstrates that Commerce does not need a preclearance regime to make a model unavailable.

-- ANNA WEBER, Berlin