White House Moves to Gatekeep Powerful AI Models

The White House has begun allowing limited access to some of the country’s most powerful artificial intelligence systems, signaling a new phase in Washington’s effort to control frontier models without shutting them down entirely.

In recent days, Anthropic was permitted to provide its advanced model, Mythos 5, to a narrow group of approved American companies and government agencies after weeks of negotiations with the Trump administration. At nearly the same time, OpenAI said it was releasing its new GPT-5.6 family only as a restricted preview for a small set of trusted partners, after a request from the White House to delay a broader rollout.

Taken together, the moves suggest that the federal government is settling on a selective-access approach to cutting-edge AI: not a blanket ban, but a system in which the most capable models may be used by approved organizations while remaining out of reach for much of the public and many overseas customers.

A Shift From Shutdowns to Gatekeeping

The latest decisions mark a notable turn from the disruption that followed a June 12 government intervention targeting Anthropic. That directive, rooted in foreign-access restrictions, barred use of certain advanced models by foreign nationals. Anthropic responded by taking Fable 5 and Mythos 5 offline globally, a sweeping step the company took as the fastest way to comply.

Now, rather than forcing a total blackout, the administration appears to be allowing a partial restoration under tighter supervision. Anthropic has been authorized to resume access to Mythos 5 for selected U.S. organizations, including some foreign-national employees connected to those approved entities.

OpenAI’s case is different in form but similar in effect. The company unveiled GPT-5.6 Sol, describing it as a next-generation system with stronger performance in coding, science and cybersecurity, along with a more advanced safety stack. But instead of making the model broadly available through ChatGPT or its application programming interface, OpenAI said it would offer only a limited preview.

That preview is being confined to a small set of partners whose participation was shared with the government, according to reporting and company statements. OpenAI has indicated that wider release is planned later, but not yet.

The Government’s New Leverage

The paired decisions amount to more than a temporary scheduling dispute between Washington and Silicon Valley. They point to a practical framework for government influence over frontier AI deployment: who gets access, under what conditions and on what timeline.

That matters because the most advanced models are increasingly seen not simply as commercial software products, but as dual-use technologies with implications for cyberoffense, biodefense, intelligence analysis and military planning. Restricting access, even temporarily, gives the government leverage while preserving availability for domestic agencies, major contractors and select companies deemed safe or strategically important.

It also creates a widening divide between what leading AI labs say they have built and what ordinary users, developers and even many enterprise customers can actually use. For months, public discussion of AI competition has revolved around benchmark scores, model releases and commercial rivalries. The latest events suggest another contest is now underway: over administrative control of deployment itself.

Why OpenAI Is Chafing

OpenAI has made clear it is unhappy with the arrangement. In describing the staged release, the company said the delay keeps its best tools from users, developers, enterprises, cyber defenders and global partners who need them.

That frustration reflects a broader tension that has been building between AI companies and policymakers. Labs want to move quickly, both to serve customers and to maintain a lead in a fiercely competitive market. Officials, especially those focused on national security, are increasingly wary of unrestricted access to systems that may be able to automate vulnerability discovery, accelerate scientific research or be adapted for misuse.

Reporting has indicated that the Office of the National Cyber Director and the White House Office of Science and Technology Policy asked OpenAI to slow-roll GPT-5.6 while the administration works on a broader testing and evaluation framework for advanced models. If that framework takes hold, the OpenAI delay may come to be seen less as a one-off and more as the template.

Anthropic as the Test Case

Anthropic has effectively become the administration’s first large-scale test case for AI access controls. The June 12 order forced the company into an abrupt global shutdown of some top models because sorting users by nationality and eligibility in real time was not operationally feasible.

The new permission for limited release suggests the government is now willing to tolerate a more calibrated system, one that restores access for vetted institutions rather than requiring a complete stoppage. But the arrangement also leaves important questions unresolved.

It remains unclear how organizations are being approved, how long the restrictions will last, whether the approved lists will expand quickly and whether similar conditions will be applied across the industry. It is also uncertain whether Anthropic’s broader Fable 5 access will be restored soon or whether OpenAI’s wider GPT-5.6 launch could face additional delays.

A Precedent in the Making

For the AI industry, the immediate significance is not just that two major companies have been slowed. It is that the White House has shown it can shape the release path of frontier systems before they reach the market at scale.

That precedent could reshape product strategy across the sector. Companies may feel pressure to design launches around phased access, government consultation and pre-vetted customer pools. Investors and enterprise customers may have to account for political approval as part of the timeline for deploying top-tier models. And foreign customers could find themselves increasingly separated from the most advanced American systems, even when those systems technically exist and are ready.

What began this month as an emergency intervention against one company is now looking more like the first implementation of a broader policy: America’s most powerful AI models may remain available, but only to the people Washington decides should have them.

Sources

Further reading and reporting used to add context:

AI News