The White House is preparing an Executive Order that would establish a voluntary pre-launch review framework for frontier AI models. According to reporting from the New York Times and The Information, the draft order would give the government up to 90 days to evaluate models before developers release them publicly. The administration hasn’t signed the order as of May 21, 2026, but multiple independent outlets, including Politico, describe it as near-final.
The 90-day window isn’t a technicality. It’s a product timeline.
For any frontier lab running quarterly release cycles, a mandatory pre-launch review window of that length restructures the entire development calendar. It means compliance work starts before the engineering work ends. Every model destined for public release would need a parallel government review track built into the roadmap.
That’s the friction industry is reacting to. According to CNN’s reporting, leading AI companies are reportedly pushing for a review window as short as 14 days, a counter-proposal that hasn’t been independently verified but, if accurate, represents a negotiating floor, not a serious operational ask. The gap between 90 days and 14 days is the gap between structural oversight and symbolic compliance.
Pre-Launch AI Review: Who Wants What
The Mythos question sits underneath all of this. Anthropic has confirmed it withheld its Mythos model from public release and established Project Glasswing, a restricted-access consortium for Mythos Preview, with up to $100M in usage credits committed to the effort. The BBC reports Anthropic describes Mythos as too dangerous to release publicly because of its hacking capabilities, a characterization that is Anthropic’s own, not independently evaluated. Still, advocacy groups have specifically cited Mythos in urging the White House to accelerate the review framework. Whether Mythos caused the EO or confirmed a direction already underway, it became the story’s anchor.
Context matters here. This EO follows the voluntary CAISI commitments labs made in 2024 and the subsequent reporting, covered previously in this hub, that the voluntary era was running out of runway. A Commerce Department announcement on model testing agreements was reportedly pulled from the NIST website in mid-May, that’s background context from prior coverage, not a new development today.
The catch is that “voluntary” does the heavy lifting in the current draft. The framework asks developers to participate, it doesn’t compel them. But voluntary frameworks with White House backing and NIST operational infrastructure have a way of becoming the floor for what comes next. The labs that participate set the standard; the labs that don’t explain why they didn’t.
Verification
Partial NYT (T2), The Information (T3), Politico (T3), Anthropic.com (T1), BBC (T2) 90-day window confirmed. 14-day industry counter is single-source (CNN), unverified. Mythos capability claims are vendor-characterization only, no independent technical evaluation.What to watch
whether the EO signs in its current form with the 90-day window intact, or whether industry lobbying narrows it before signature. The specific review window in the final signed text is the compliance planning variable. Until it’s signed, any internal pre-launch review program should be designed to flex between 14 and 90 days, building for the government’s ask while acknowledging the final number isn’t confirmed.
The real question is what “voluntary” means once your competitors have all signed up. Frontier labs that participate in the review framework gain a government imprimatur that non-participants won’t have. That asymmetry may do more to drive adoption than any enforcement mechanism.