A June executive order gives the federal government up to 30 days to assess the most capable AI systems for cybersecurity and national security risk before public release. The two largest U.S. labs are already complying, even as they call it a temporary arrangement.
OpenAI and Anthropic, the two companies building the most capable AI systems in the United States, have begun letting the Trump administration preview their newest models before the public can use them. Both companies are calling the arrangement temporary. Both have already reshaped their release plans around it.
On June 26, OpenAI said it is limiting release of GPT-5.6 Sol, its newest flagship model, to a small group of customers approved by the White House. The model is being held back during what the company describes as a 30-day cybersecurity and national-security review. The same day, Anthropic said the administration had cleared a limited release of Mythos 5, its cybersecurity-focused model, after roughly two weeks of the company effectively holding it back. Anthropic is also restricting Mythos 5 to cyber defenders and critical infrastructure providers.
The reviews are happening under an executive order President Trump signed on June 5, titled "Promoting Advanced Artificial Intelligence Innovation and Security." The order sets up what the White House describes as a voluntary federal pre-release assessment process for advanced AI systems, with a window of up to 30 days before a model can be made broadly available. The order does not define "trusted partners," a term the AI labs have adopted themselves, and participation is, on paper, optional.
In practice, the two largest U.S. AI labs are already inside the process. The Associated Press, the Guardian, the Washington Post, and Wired independently described the moment as the first time the U.S. executive branch has previewed frontier commercial AI before public release. ABC News and the trade outlet SecurityWeek confirmed the same pattern across both labs.
OpenAI and Anthropic are presenting their participation as a one-off accommodation, not a permanent gate. In its own announcement, OpenAI said it does not believe "this kind of government access process should become the long-term default." Anthropic, in parallel, framed its Mythos 5 release as a narrow deployment to specific defenders rather than a general release. The framing is consistent with the executive order's own voluntary language.
But the operational picture already looks compulsory at the top of the market. OpenAI's own deployment safety page rates GPT-5.6 Sol at "High" capability for cybersecurity and biological and chemical risk under the company's Preparedness Framework, the internal threshold the company set for phased release. In other words, the kind of model the government now wants to preview is the same kind of model OpenAI's own safety system was already flagging for staged rollout. The two gates are converging.
Anthropic's path to June 26 also tells a longer story. The company had earlier withdrawn Mythos 5 and a sibling model called Fable 5 after complying with a separate Trump administration directive restricting their use by foreign nationals. The cybersecurity-focused Mythos 5 has now been approved for limited U.S. release, but only after weeks of what was, in effect, a hold.
Several structural questions are now open. The executive order does not name a statutory category of "trusted partner"; the term is being defined by the companies and the White House in practice. Smaller AI labs, academic researchers, and foreign-domiciled customers do not yet have a clear path to the same models. The 30-day window can, in principle, stretch. The order's "voluntary" label could harden into precedent for the next administration, or quietly lapse once attention moves elsewhere.
For now, the practical answer is simple. The newest AI from the two biggest U.S. labs is no longer sold the way consumer software is sold. It is sold the way defense suppliers sell to the government, with the government as the first customer and the public waiting until the government says otherwise.