The government now gets a look at the most powerful AI models before the public does

For twelve days this month, the most capable model OpenAI had built sat behind a government gate. GPT-5.6 was ready on June 26, but the company held it to about 20 vetted organizations at the request of two White House offices, the Office of the National Cyber Director and the Office of Science and Technology Policy, and released it broadly only on July 9. The White House is careful to say it gave no approval, and that release decisions rest with the companies. Still, the most anticipated US model of the year waited on Washington before it reached the public.

The pause followed Executive Order 14409, signed June 2. It does something narrower than the headlines suggest: the government is not approving AI. The order creates a category, the “covered frontier model,” and calls for a voluntary framework through which labs could give federal reviewers access to such a model before releasing it to other trusted partners.

The order calls for a classified benchmark, with the NSA deciding which models it covers, built to test one thing above all: a model’s advanced cyber capabilities. GPT-5.6 was a major release after the order, and it came before the framework was finished.

Two words carry most of the weight, and both are unsettled. The first is “covered.” The order doesn’t say, publicly, how capable a model must be to qualify. That threshold is what the classified NSA benchmark is meant to set, and it’s due by August 1, the deadline the order gave agencies to finalize the process and publish the rules. Until then, “covered” means what the reviewers decide, model by model.

The second word is “voluntary.” No company is required to submit anything. But opting out risks federal cloud contracts, government AI procurement, and the light regulatory touch the industry has enjoyed so far. For OpenAI, Google, and Anthropic, the government is not a customer you walk away from. That is how the framework wins compliance without a statute.

The review is the gentle version. There is a blunt one, too. On June 12, three days after Anthropic released Claude Fable 5, the Commerce Department ordered the company to switch off Fable 5 and its most capable model, Mythos 5, for every foreign user in the world, including Anthropic’s own foreign-national staff, after a jailbroken safeguard was flagged to Washington by Amazon. The models stayed dark for 19 days, until Commerce lifted the order on June 30. That was a mandatory export-control order, not a voluntary review, but it ran through the same office.

Commerce’s order on Fable 5 and Mythos 5 was the first time export controls had been used to restrict foreign access to frontier AI models, according to one law-firm analysis.

None of this is settled. The threshold is undefined, the formal rules aren’t published, and “voluntary” could harden or soften with the next fight over mandatory AI regulation. But the direction is clear enough to plan around: for the top tier of AI, the US government has placed itself between the lab and the launch, and it is using procurement, not law, to make that stick.

So what: for the corridor, the catch is foreign access. Separately, Washington has already used export controls to restrict foreign access to frontier models, which would put Gulf buyers downstream of a US decision. The opening that creates is small but real: non-US model-makers like Canada’s Cohere.

Related

• US grants UAE A:5 status, ending license rules on advanced AI chips — the hardware side of the same corridor valve
• Cohere and Saudi Arabia’s Humain partner on 50MW sovereign AI compute deal — a non-US model-maker as the un-gated alternative