OpenAI's Strongest AI Ever Just Got Restricted by the Government

Two of the most powerful AI models ever built just launched, and almost nobody is allowed to use either one. That is not an accident. That is the government.

What OpenAI previewed

On June 26, OpenAI previewed GPT-5.6, a family of three new models. The flagship is a monster on paper. It edges out Anthropic's Claude Mythos 5 on coding benchmarks, and on cybersecurity testing it is competitive with Mythos while using about a third of the compute.

Here is the part that should get your attention: almost nobody can use it. The flagship is locked behind a preview available to roughly 20 vetted organizations, hand-picked, with their names shared with the US government before the model even launched.

Why it is locked down

The reason is that it is very good at finding security vulnerabilities. In testing against real, hardened software, including Chrome and Firefox, it found actual bugs and pieces of working exploits. It did not chain them into a full attack, but it got close enough that OpenAI and the government both decided the general public does not get this one yet.

If that sounds familiar, it should. The same month, a similar restriction hit Anthropic's Claude Mythos. Two competing labs, two flagship models, both throttled by the same government process within a few weeks.

The government as checkpoint

So let's be clear about what is happening. This is not one company being singled out. It is the US government positioning itself as the checkpoint between the public and the most capable AI models on the planet, under a new executive order that creates a category called covered frontier models and gives federal officials the power to evaluate what these systems can do before the public ever touches them.

Is that a bad thing? I do not think it is that simple. A model that can independently find real vulnerabilities in Chrome and Firefox is not a toy. If that capability lands in the wrong hands before defenders get a chance to patch what it finds, that is a real problem, not a hypothetical one. There is a legitimate case for slowing it down.

But here is the part worth sitting with. OpenAI itself essentially said the quiet part out loud: it does not think this kind of government gatekeeping should become the permanent default, because it keeps these tools away from the developers, defenders, and businesses who actually need them. And they have a point. When the two most advanced AI models on the planet both require a phone call to Washington before you can get near them, that stops looking like a safety policy and starts looking like a choke point. Choke points have a way of outliving the promise that they are temporary.

What comes next

OpenAI says the flagship will open up more broadly in the next few weeks. Pricing is set at $5 per million input tokens and $30 per million output tokens, the same as its last flagship. The middle model runs at about half that, and the smallest is the cheap, fast option most people will actually use day to day.

My take

Watch this space, not because of what the flagship can do, but because of what the pattern says about where AI is heading in 2026. Two labs, two flagship models, one government checkpoint, in the same month. That is not a coincidence anymore. That is a policy. Whether you see it as smart caution or a slippery slope, it will shape which AI companies win and which get stuck waiting on a preview list. This is not the last time we will be talking about who gets to hold the keys.