The White House is racing to finalize a set of voluntary standards governing how the most powerful AI systems get tested and released, and the announcement could land as early as the week of July 7, according to reports on the negotiations. The framework, hammered out in a rush of meetings with OpenAI, Google and Anthropic, would for the first time give the U.S. government a formal role in reviewing frontier models before they reach the public.
The push follows a striking precedent set two weeks ago, when the Trump administration asked OpenAI to hold back the wide release of GPT-5.6 and instead let a small set of government-approved partners in first. It marked the first time Washington had preemptively asked a U.S. AI company to restrict a launch, and it turned an abstract policy debate into a live test of how far the government can shape release decisions at the labs.
What the standards would cover
The proposed rules would establish benchmarks for what counts as a frontier model and set clearer timelines for testing and deployment, according to reporting on the talks. They are also expected to define who can access the most capable systems, both domestically and abroad.
Technical teams from the AI firms have met repeatedly with White House officials over the past week to reach a final agreement, per those reports. The sticking points are the ones that matter most in practice: how long the government's review window should be, and the precise threshold at which a model is deemed "frontier" and therefore subject to the process. A person familiar with the talks said the Commerce Department's Center for AI Standards and Innovation (CAISI) and the National Security Agency are expected to play central roles in setting and monitoring the standards.
The framework grows directly out of an executive order President Trump signed June 2, titled "Promoting Advanced Artificial Intelligence Innovation and Security." That order directed agencies to build a voluntary protocol through which developers could submit models for evaluation. Under the order, once a system is designated a "covered frontier model," a developer may give the government access for up to 30 days before releasing it more broadly, subject to confidentiality, cybersecurity and intellectual-property protections. The order tasked the departments of Treasury, War and Homeland Security with designing that framework within 60 days, putting deliverables on a track toward an August 1 deadline. The NSA and the Cybersecurity and Infrastructure Security Agency were directed to build a classified benchmarking process to flag models with advanced cyber capabilities.
The GPT-5.6 flashpoint
The urgency traces to GPT-5.6. The White House's Office of the National Cyber Director and Office of Science and Technology Policy asked OpenAI to limit the rollout as the administration builds out its testing regime, according to a source who spoke to Axios. The government, that source said, intervened because GPT-5.6 has "Mythos-like" capability, referring to Anthropic's frontier model, not because officials had suddenly decided to take a heavier hand. "This is what's happening with models of that caliber," the source said.
OpenAI CEO Sam Altman laid out the arrangement in a memo to staff, first reported by The Information. He said the government would approve access "customer by customer" during a preview period, and that if the limited release went smoothly, OpenAI hoped to follow with a broader launch a "couple of weeks later." Altman was blunt that the setup should not become permanent.
"We've made clear to the U.S. government that this is not our preferred long term model, and will work with them and others in industry to achieve a more sustainable approach for future releases," Altman wrote, according to The Information.
Google, for its part, is also in discussions with the government as it prepares advanced coding models with stronger cyber capabilities than earlier versions, according to reporting on the talks. The moment arrives just as Anthropic and Washington reached a truce of their own: the Commerce Department this month lifted export restrictions on Anthropic's Fable 5 and Mythos 5 models, less than three weeks after cutting off access over national-security concerns.
Why It Matters
For most of the modern AI era, U.S. labs decided on their own when a model was ready to ship. These standards would insert the government into that decision at the frontier, a soft form of gating built on benchmarks, review windows and access controls rather than outright approval.
The framing is national security. Officials and executives alike have grown wary of what happens when nation-state hackers, cybercriminals or rogue insiders get hold of models with strong offensive cyber capabilities. That concern is what elevated GPT-5.6 and Mythos-class systems from product launches to matters of state.
But the "voluntary" label carries real weight. Participation is not mandatory, and developers are not legally required to engage the government before releasing a model. That leaves the regime resting on cooperation and leverage rather than law. The internal fight over how restrictive and mandatory the underlying executive order should be delayed its signing by weeks, a sign of how contested the line between voluntary and compulsory remains. Labs are also caught in a bind: slow-walking releases to satisfy Washington while racing increasingly capable Chinese open-source models that answer to no such review.
What to Watch
Whether the standards are unveiled the week of July 7, as reported, and how firmly the "frontier" threshold and review timeline are drawn — the details will determine how many models actually get caught by the process. Watch whether GPT-5.6 gets its broader release on Altman's "couple of weeks" timeline, which will be the first real test of the staggered model. And watch the pressure toward something more binding: Altman has already signaled the current arrangement is unsustainable, and the gap between a voluntary framework and mandatory rules is the fault line this whole effort is built on.
"We've made clear to the U.S. government that this is not our preferred long term model, and will work with them and others in industry to achieve a more sustainable approach for future releases."— Sam Altman, CEO, OpenAI