The 90-Day Gap: Who Stands Where on Pre-Launch AI Review and What the Negotiation Reveals

Three news outlets confirmed it independently.

The New York Times, The Information, and Politico each reported, from what appear to be independent sourcing, that the White House is finalizing an Executive Order establishing a voluntary pre-launch review framework for frontier AI models. The Information added the specific operational detail that matters most: AI labs would share models with the government up to 90 days before public release. That’s the number compliance teams need to build around.

But “voluntary” and “90 days” are doing very different kinds of work here. Understanding why requires mapping each stakeholder and what they’re actually protecting.

What’s Confirmed and What Isn’t

Before the map, the evidence baseline. Three claims are confirmed. Three require qualified framing.

Confirmed: The EO exists in near-final draft form, per NYT, The Information, and Politico independently. Confirmed: The government’s proposed review window is 90 days pre-launch, per The Information’s reporting. Confirmed: Anthropic withheld its Mythos model from public release and established Project Glasswing as the restricted access consortium for Mythos Preview, committing up to $100M in usage credits – per Anthropic’s own published announcement.

Qualified: CNN reports industry participants are pushing for a review window as short as 14 days. That figure is single-source and couldn’t be independently verified in this reporting cycle, treat it as a reported negotiating position, not a confirmed counter-proposal. Qualified: Anthropic’s Mythos model is described as possessing autonomous cybersecurity exploitation capabilities. That characterization comes from Anthropic and has been reported by the BBC as Anthropic’s own stated rationale for the withholding. No independent technical evaluation confirms the specific capability level. Qualified: Sources suggest the EO could be issued imminently. A signing date hasn’t been officially confirmed.

The distinction matters because the 14-day counter-offer and the specific capability characterization are the two most-cited data points in downstream coverage, and neither is independently verified.

The Stakeholder Map

Five actors have documented positions. Their interests don’t line up cleanly.

The White House. The administration’s case for a 90-day window is a national security case, not a safety culture case. The framing from Politico describes the directive as asking companies to submit models for review by federal agencies, plural. That signals NIST isn’t the only institution in the review architecture. Defense and intelligence agencies presumably want the runway to assess cybersecurity and dual-use risks before a model enters the commercial ecosystem. Ninety days gives technical evaluators time to run meaningful assessments, not just cursory checks.

NIST. The Commerce Department’s role is operational, NIST would presumably run or coordinate the review infrastructure. This hub covered in May that NIST had reportedly pulled an announcement about model testing agreements from its website. That’s prior context, not a new development, but it signals that the institutional machinery for this review program has been in motion for weeks. The agency that built the AI Risk Management Framework is the natural home for a pre-launch evaluation process.

OpenAI and Anthropic. The labs aren’t monolithic on this. The Information’s reporting indicates OpenAI is actively engaged in the drafting conversations. Anthropic’s position is more complex: it voluntarily built the architecture the EO is codifying, restricted access, government partnership, controlled capability release, but its Project Glasswing model is substantially more restrictive than what the EO proposes. Anthropic already chose 0-day public release for Mythos, not 90-day. The EO would actually represent a loosening of what Anthropic has voluntarily adopted for its most capable model.

AI Safety Advocacy Groups. The advocacy community has been specific in citing Mythos as the reason mandatory oversight is urgent. Their public statements don’t just support the EO in general, they name the model and argue its capabilities justify government screening. That specificity has amplified the pressure on the White House to move quickly. Whether Mythos caused the EO or accelerated a process already underway, it became the symbolic anchor.

Frontier Labs Collectively. The reported 14-day counter-offer, if accurate, tells you the labs understand the EO is coming and are negotiating the terms, not the principle. A 14-day window doesn’t give evaluators time to do anything meaningful. Its practical function would be notification, not assessment. The labs know this. What they’re protecting isn’t the ability to avoid review, it’s the ability to control the calendar and prevent a competitor from using the review process to gain advance intelligence on unreleased capabilities.

Why the Gap Is Larger Than It Looks

The distance between 90 days and 14 days isn’t about review duration. It’s about who controls the development calendar.

At 90 days, a pre-launch review process becomes a structural constraint on release strategy. Labs operating on quarterly cycles would need to freeze release candidates 90 days before their target date, meaning engineering and safety work needs to complete five to six months before the public sees the model. That’s a meaningful reorganization of internal timelines, not a compliance checkbox.

At 14 days, the review is informational. Government reviewers get advance notice and a look at what’s coming. They don’t have the time to run independent evaluations, red team the model, or coordinate multi-agency assessment. The 14-day window is the version of oversight that doesn’t actually constrain anyone.

The real question is whether a voluntary framework can hold the 90-day line. Labs that participate in the framework under a 90-day window set the operational standard. Labs that negotiate down to 14 days, or don’t participate, create an asymmetry where the most cooperative labs have the most constrained timelines. That’s an incentive structure that rewards non-participation, which is why the “voluntary” framing in the current draft is the detail worth scrutinizing most closely.

What the Mythos Architecture Tells Compliance Teams

Anthropic’s Project Glasswing is the closest existing model of what restricted frontier AI access can look like in practice. The restricted consortium includes a commitment of up to $100M in usage credits for Mythos Preview. Access is controlled. The model doesn’t go to the general public.

That structure didn’t emerge from government mandate, it emerged from Anthropic’s own assessment that Mythos’s capabilities warranted it. The CAISI voluntary commitments that labs made in 2024 established a precedent for self-imposed governance. What the EO does is take that precedent and build a government review layer on top of it, creating a process where the government gets the evaluation window before the lab makes its release decision.

For compliance teams at frontier AI developers, the practical planning question is this: if a 90-day pre-launch review becomes the framework, voluntary or mandatory, what does an internal pre-launch compliance program need to look like to meet that standard? The NIST AI Risk Management Framework provides the closest existing scaffold. Article 9 of the EU AI Act’s conformity assessment requirements for high-risk systems offers a parallel structure from a different jurisdiction. Neither maps perfectly, but both give compliance architects a starting structure.

What Comes Next

The EO signing is the first marker. Its text, specifically whether the 90-day window survives the negotiation, is the compliance planning input. A signed EO with a 90-day window starts a clock for labs to build the internal infrastructure to comply. A signed EO with a shortened window narrows the government’s practical review capacity and signals that the mandatory version, when it comes, will need to specify the timeline more precisely.

Don’t expect this to stay voluntary. The voluntary framework is structurally positioned as the instrument that proves the concept. If participating labs find it manageable and government evaluators find it useful, the argument for making it mandatory strengthens considerably. The labs that build functional 90-day pre-launch compliance programs now will have a significant advantage when the mandatory version arrives, their process already exists.

The precedent being set in the next 30 days will define what pre-launch AI oversight means for the next several years.