Two weeks ago an export-control letter pulled Anthropic's Fable 5 and Mythos 5 offline for the entire planet in a single afternoon. This week OpenAI got the polite version of the same power. Sam Altman told staff that GPT-5.6 won't ship the way its predecessors did. It goes out first as a limited preview to a short list of partners, and during that window the government approves access customer by customer.

So the administration didn't reach for the off switch this time. It reached for a guest list. The Verge called it a more favorable deal than the one Washington handed Anthropic, and that's true on its face. Nobody's product vanished. No foreign-national ban forced a global shutdown just to stay compliant. OpenAI seems to have walked into the arrangement willingly, having spent weeks working with the Office of the National Cyber Director and the Office of Science and Technology Policy on the rollout before the request ever arrived.

The safety argument underneath it isn't stupid, which is what makes this hard. Both OpenAI and the administration reportedly judge GPT-5.6 to be on par with Mythos, the model that started this whole mess. If a system really sits at that level of capability, shipping it to everyone on day one is the reckless option, not the brave one. A vetting period, a small group of trusted users while you watch for trouble, is the kind of staged caution frontier-safety researchers were asking for long before any government got involved. Judge the staggered launch on its own and it looks more like responsibility than overreach.

The precedent is a different animal. The caution isn't the problem; the mechanism doing the cautioning is. Access decided one customer at a time, by the executive, on security judgements it never has to publish, with no statute written down anywhere. Axios noted this is the first time Washington has preemptively told an American company to hold back a launch, and that "first time" is the whole story. It's the same discretionary power that darkened Fable, used with a lighter hand. The lighter hand is what makes it dangerous. A shutdown provokes lawyers and headlines; a favor provokes gratitude, and gratitude is far easier to extend.

I called the last version of this a license nobody wrote, and that's exactly what arrived, only now the company is holding the door open. Altman told staff this isn't OpenAI's preferred long-term model. I believe him. The trouble is that long-term models are built out of short-term exceptions, and the exception here is that the government now signs off on who gets to use a product before the public ever sees it. Do that twice and it stops being an exception. It becomes the way frontier models ship.

Sources: