On Friday, OpenAI confirmed that its newest model, GPT-5.6 Sol, will be released only to “a small group of trusted partners” pre-approved by the Trump administration. The Information first reported the plan Thursday; Axios followed with a source confirming the White House had formally asked OpenAI to limit the rollout. By Friday morning, the company had complied.

The framing, carefully laundered through the administration’s preferred outlets, was security. GPT-5.6’s capabilities are so advanced, the argument goes, that a broader release would pose an unacceptable cybersecurity risk. The model is reportedly comparable to Anthropic’s Mythos, which was pulled from public access just weeks ago under the administration’s new export-control order.

This is tidy. It is also a dodge.

The real story isn’t that the government wants to keep a dangerous model out of the wrong hands. The real story is that the government just got to decide which hands are right — and the company it picked for the velvet-rope treatment happens to be the one that has spent the last two years embedding itself in Washington like a tick on a deer.

The Mythos Precedent Was the Template, Not the Exception

Rewind to May 2026. The administration’s executive order on AI oversight — delayed just hours before signing, as the Post reported on May 21, then quietly revived — gave the Commerce Department sweeping authority to classify advanced AI models as dual-use technologies subject to export controls. Anthropic’s Mythos and Fable were the first casualties, yanked from public access over fears about what a sufficiently capable model might enable in the wrong jurisdiction.

The market read this as even-handed. Two models, both frontier-class, both restricted. Symmetry.

But symmetry is the wrong frame. The government’s power to restrict a model is not the same as the government’s power to designate a model’s approved user base. One is a blunt tool. The other is a licensing regime — and licensing regimes are, historically, where competition goes to die.

Who Gets to Be a ‘Trusted Partner’?

The administration has not published a list of approved users for GPT-5.6 Sol. A source familiar with the matter told Axios the group would be “small.” OpenAI’s own language — “trusted partners” — is the kind of phrase that does a lot of work while meaning nothing at all.

But the pattern is already visible if you squint. OpenAI has spent the past two years aggressively courting defense and intelligence contracts. It has hired former National Security Council staffers. Sam Altman has made multiple trips to Mar-a-Lago. The company’s government affairs operation is, by any measure, the best-funded and most aggressive in the AI industry.

Anthropic, by contrast, took a different route. It emphasized safety research and constitutional AI. It published its model cards. It built relationships with academics and civil-society groups. What it did not do — at least not at the same scale — was build a political machine.

Now Anthropic’s models are locked down, and OpenAI’s are rolling out to a hand-picked guest list. If you believe this is a coincidence, I have a Defense production contract to sell you.

The Cybersecurity Argument Collapses Under Its Own Weight

If GPT-5.6 Sol genuinely poses a cybersecurity risk serious enough to warrant government vetting of every end user, then restricting it to a small group does not solve the problem — it concentrates it. A model capable of generating novel zero-day exploits is not made safer by giving it exclusively to organizations that are themselves high-value targets for espionage.

“The idea that a model is too dangerous for the public but safe enough for a few defense contractors is self-evidently absurd,” one security researcher at a major cloud provider told me in a Slack DM on Friday. “Either the capability is manageable or it isn’t. Picking who gets to hold the grenade doesn’t defuse it.”

The administration’s real concern has never been the model. It has been the distribution channel. A model restricted to government-approved partners is a model whose economic rents flow through a valve the government controls. That is not a security policy. That is an allocation mechanism.

Industrial Policy Has Become Favoritism With a Flag Pin

The Trump administration has been explicit about its desire to create national champions in AI. That is a defensible goal. What is less defensible is the quiet slide from “we will support American companies” to “we will pick which American company wins.”

The GPT-5.6 restriction marks the moment that slide became official. The government is no longer just funding research or setting safety standards. It is deciding who gets to ship product — and the company that gets to ship first is the one that played the access game best.

Google’s DeepMind division is watching this from London. Meta’s AI research group is watching from Menlo Park. Every startup that was hoping to compete on model quality rather than political relationships is running the numbers on how fast it can open a D.C. office.

The market will internalize the lesson quickly: technical capability matters less than regulatory proximity. That is not a recipe for American AI dominance. It is a recipe for an industry where the most important engineering decisions happen in a congressional office building, not a server cluster.

The irony is that OpenAI’s model may well be excellent. The engineers who built GPT-5.6 Sol almost certainly earned their achievement on the merits. But nobody will ever know for sure, because the terms of its release now carry an asterisk that no benchmark can remove: this model didn’t just pass safety tests. It passed a political one.

Sources