The Trump administration has asked OpenAI not to release its next model, GPT-5.6, broadly, but rather as a limited preview. The government is approving access on a customer-by-customer basis, as reported by The Information. CEO Sam Altman shared the arrangement with staff during an internal Q&A. The request came from two agencies: the Office of the National Cyber Director and the Office of Science and Technology Policy.

The practical effect is clear. OpenAI's release pace is now partially running on Washington's schedule. Instead of a simultaneous rollout, there is restricted access for a small group of partners while agencies complete their review.

The trigger can be found in OpenAI's own Safety Card. For the first time, all three GPT-5.6 variants simultaneously reached the High tier for biological and cyber capabilities. It was precisely this threshold that triggered the government review. The self-reasoning control rate of Sol tripled from 0.4 percent in GPT-5.5 to 1.3 percent. OpenAI has marked this trend as under investigation, not resolved.

This is, de facto, AI regulation. And it is new. These models do not hit the highest risk tier under OpenAI's own framework, yet the authorities are intervening regardless. Former White House advisor Dean Ball already calls the voluntary submission process a de facto involuntary licensing regime, entirely without defined standards or a formal procedure.

There is an uncomfortable context here. Anthropic has accused Alibaba of brazenly and illegally siphoning off AI capabilities in a distillation attack. According to Bloomberg, actors linked to Alibaba allegedly conducted 28.8 million exchanges with Claude models via roughly 25,000 fake accounts between April 22 and June 5. Anthropic outlined this in a letter to the Senate Banking Committee.

This is where allegations of regulatory capture arise. Investor Bill Gurley put it sharply: Anthropic could have sued in court, as others do. Instead, they want something far more valuable—government protection from competition for years to come. A court cannot provide that. Regulation can.

And Gurley followed up: If your models are supposedly close to AGI and are the world's most dangerous for cyberattacks and cyberdefense, why don't they detect distillation themselves in real time? Why write letters to Washington instead of deploying your own cutting-edge technology? A fair question.

OpenAI itself does not view this arrangement as its preferred model. Its blog post explicitly states that they do not believe this type of government gating should become the long-term standard. It keeps the best tools away from users, developers, enterprises, and global partners. They describe it as a short-term step toward broader availability in the coming weeks.

The word short-term is crucial. OpenAI is treating this as a temporary concession and is working with the administration on a framework for future models. This leaves an open question: whether this customer-by-customer hurdle also applies to international customers and whether it will be extended to all future releases.

There is a silver lining. Anyone looking to keep up while the frontier is walled off has an option: open source. Sovereign AI strategies will likely be built on open weights. That makes open models more important than ever. They lag behind the frontier today, but they are accessible. And they can be run locally.