If you were expecting OpenAI’s usual playbook for GPT-5.6 — big announcement, API goes live, everyone piles in — you can forget about it. The Trump administration has told OpenAI to stagger the release. No broad public launch. Instead, access gets approved customer by customer, with government sign-off.
What We Know
Sam Altman reportedly told staff that during a preview period, the government would approve access customer by customer. Selected partners go first. If that limited release goes well, broader availability follows a couple of weeks later.
OpenAI employees apparently worked closely with the government on the rollout plan. This isn’t a case of Washington springing a surprise — it’s a coordinated process.
Who’s Calling the Shots?
Two offices in the White House are leading the effort: the Office of the National Cyber Director and the Office of Science and Technology Policy. The legal backbone is an executive order Trump signed on June 2, directing AI companies to submit frontier models for government testing before any public release.
That’s not a suggestion. It’s a directive.
Why This, Why Now?
The concern is cybersecurity. Frontier models — the most capable models from any given lab — can identify and exploit software vulnerabilities at a speed that simply didn’t exist before. We’ve seen that clearly with Anthropic’s Project Glasswing, where Mythos found over 10,000 critical vulnerabilities in a single month.
If a model can do that defensively, the logic goes, a similarly capable model could do it offensively. And that’s what keeps Washington up at night.
Anthropic Has Been Doing This Voluntarily
Here’s the interesting parallel: Anthropic has been restricting access to Mythos voluntarily for months through Project Glasswing. Same logic — equip defenders first, then widen access. But Anthropic made that call on its own. With OpenAI, the instruction came from the government.
That distinction matters. One is a company choosing caution. The other is the state requiring it.
My Take
This is a genuinely new moment. For the first time, the US government is directly intervening in how a frontier AI model gets released — not through legislation, but through executive order and direct communication. Whether you see that as responsible oversight or government overreach probably depends on where you sit.
But the signal to the industry is unmistakable: the era of just shipping your most powerful model to the world is over. And honestly, given what Mythos found inside classified US systems in a matter of hours, I can understand the caution. The real question is whether ‘customer-by-customer approval’ is a workable mechanism — or whether it quickly turns into a political gatekeeping problem where access depends on who you know in Washington rather than what you need the model for.
Sources: TechCrunch: The White House is asking OpenAI to slow the release of GPT-5.6, CNBC: Trump administration told OpenAI to share GPT-5.6 with select partners first