The Trump administration just asked OpenAI to stagger the release of GPT-5.6 over national security concerns. Sam Altman told staff on Wednesday that the government will be approving access to the new model "customer by customer" during a preview period before any broader public release. This is a direct consequence of what happened with Anthropic weeks ago. When Anthropic released its Mythos and Fable models, the Commerce Department slapped export controls on them, effectively pulling the most powerful versions offline. Anthropic is still in negotiations with Commerce Secretary Howard Lutnick to get those models fully restored. A company has even filed a lawsuit arguing the government's use of export controls against AI models exceeds the bounds of the Export Control Reform Act. OpenAI appears to be taking the cooperative route to avoid the same fate. Rather than release GPT-5.6 broadly and risk getting hit with restrictions after the fact, the company is working with the government upfront. Altman framed it as the fastest path to a wide release, saying he hopes a general rollout could follow "a couple of weeks later" if the preview period goes well. What's emerging is a de facto licensing regime for frontier AI models. The government is not passing legislation that says it can approve or deny access to AI models. There is no formal regulatory framework for this. Instead, it is using a combination of export controls, voluntary safety review agreements, and direct pressure on companies to achieve the same result. OpenAI, Anthropic, Google DeepMind, Microsoft, and xAI have all agreed to provide the government early access to new models for national security evaluations. The conflict of interest dimension is also worth noting. Commerce Secretary Lutnick, who is leading the crackdown, has reported financial ties to OpenAI, Anthropic's chief competitor. Congressional members have raised this publicly alongside concerns about the precedent being set. This is the new reality for AI development in the United States. The most powerful models now require government sign-off before they reach the public. No law was passed. No vote was taken. The government now decides who gets access to the most powerful AI models and who doesn't. That is not national security. That is a licensing regime built through enforcement actions and backroom pressure.