A June cybersecurity executive order set up a 30 day pre release process that the AI labs publicly call temporary, but the rollout of their newest models is already being gated by it.
The US government is now sitting in the release path of the most powerful AI systems built in the United States. Not through legislation. Through a cybersecurity review framework that two of the country's most consequential AI laboratories have concluded they cannot refuse.
On June 26, 2026, OpenAI began a limited preview of its next-generation GPT-5.6 series, headlined by a model called Sol, with two sibling variants called Terra and Luna. Access was restricted to a small group of "trusted partners," and the company shared that partner list with the federal government. The framing from OpenAI was explicit: this is a temporary arrangement while the Trump administration builds out a broader "repeatable process" for reviewing frontier AI, and the company does not want it to become permanent. "We don't believe this kind of government access process should become the long-term default," OpenAI said, framing the preview as a short-term step on the way to broader availability "in the coming weeks."
That phrasing matters because Anthropic, the other US frontier lab with a public model release in the pipeline, had been running into the same wall. Anthropic's newest system, Mythos 5, had been effectively blocked for about two weeks after the Trump administration directed restrictions on use by foreign nationals. On June 26 the restrictions were partially lifted, according to CNBC's coverage of the broader review framework, allowing limited redeployment to a narrow set of "cyber defenders and infrastructure providers." The trigger sits in a June executive order signed by President Donald Trump, which asks certain AI companies to voluntarily submit their most advanced models for government review up to 30 days before public release.
Read those two stories together and the picture sharpens. The labs are not choosing to delay their own launches. They are complying with a process that has no statutory basis and no published review criteria. It is voluntary in the same way an invitation from the most powerful customer in your market is voluntary.
Dean Ball, a former White House AI adviser, has called the framework exactly that in writing: a "de facto involuntary licensing regime," per the original reporting. That is the operational story underneath the model names.
What changes from here depends on three open design choices the executive order leaves undefined.
First, what counts as a "trusted partner." OpenAI has not published the list. Anthropic's redeployment is limited to cyber defenders and infrastructure providers, which is a narrower category with operational meaning, but still no published criteria. If the trusted-partner tier quietly becomes the de facto US distribution channel for frontier AI, the most capable class of AI systems in active development, the difference between "limited preview" and "release" stops meaning much.
Second, what the 30-day review window actually reviews. The order frames this as cybersecurity review, with an emphasis on cyber and biosecurity risk in frontier models. OpenAI's GPT-5.6 Sol benchmarks, published with the preview, were selected to show defensive strength: state-of-the-art on Terminal-Bench 2.1, a public coding benchmark; stronger biological reasoning on GeneBench v1 than its predecessor; and cyber-offense performance on ExploitBench squared that is competitive with Anthropic's Mythos preview at roughly one-third the output tokens. OpenAI also highlighted new "max" reasoning effort and "ultra" subagent mode in Sol, with safety guardrails built into the model's core behavior rather than as a separate classifier-style filter. Anthropic has not published equivalent numbers for Mythos 5 in the same window. Whether the review window examines model capability against national-security thresholds, deployment safety plans, or both, is not on the record.
Third, what happens if a lab says no. The order is voluntary. The pattern so far is that when the Trump administration directed restrictions on use by foreign nationals roughly two weeks before the OpenAI preview, Anthropic complied and withdrew both Fable 5 and Mythos 5. Fable 5 had been Anthropic's most powerful public model at the time. The cost of refusal is not theoretical.
The wider concern, which the labs themselves are starting to surface, is that undefined standards produce endless launch delays, hand a structural advantage to Chinese competitors racing to ship comparable systems, and put at risk the billions of dollars in AI infrastructure buildouts already underway. OpenAI's own framing, preview now and broader availability "in the coming weeks," implicitly concedes the delay is being absorbed rather than chosen.
What to watch next: whether the Trump administration publishes criteria for who counts as a trusted partner, whether the 30-day review window ever closes into a defined regime, and whether Congress treats the framework as a fait accompli or as the first draft of legislation. OpenAI's preview is publicly framed as a weeks-long step. If "weeks" stretches without a published review process, the de facto regime has arrived without anyone voting on it.