Gated Intelligence

Two weeks ago I wrote about Anthropic pulling Fable 5 and Mythos 5 offline overnight because of a US government directive. The models vanished. No migration window, no deprecation notice, just gone.

This week the pattern continued, and it got more formal. OpenAI released GPT-5.6, their newest model family, but only to customers approved by the US administration [1]. Not the whole internet. Not all paying API customers. About 20 vetted organizations, hand-picked during a preview period that the White House oversees.

The same day, Anthropic announced that the government lifted restrictions on Mythos 5, allowing redeployment to a small group of cyber defenders and infrastructure providers [2]. So the model that was dangerous enough to switch off on June 12 is now safe enough for selected use on June 27, under conditions neither Anthropic nor the government have fully explained.

This is not a bug. This is the new framework.

What GPT-5.6 actually is

OpenAI launched three models: Sol (flagship, deep reasoning), Terra (mid-tier, high volume), and Luna (fast, affordable, everyday). Sol gets two extra modes: Max for extended reasoning and Ultra for multi-agent problem solving [3]. Pricing starts at $5 per million input tokens for Sol, roughly half of what Anthropic charges for Fable 5 [4].

OpenAI says the model is especially strong at coding, cybersecurity, and biology. They dedicated 700'000 A100e GPU hours to automated red-teaming before launch. They also revised their preparedness framework in April, removing some areas of previous study, which is worth noting [5].

The capabilities are not the story here. The story is who gets to use them.

The gating mechanism

Earlier in June, the US president signed an executive order on AI oversight that lets the federal government vet frontier AI systems for national security risks for up to 30 days before public release [6]. Participation is technically voluntary. The framework is not fully developed. But both OpenAI and Anthropic are complying, because the alternative is presumably worse.

OpenAI was blunt about not wanting this to be permanent: "We don't believe this kind of government access process should become the long-term default," the company wrote. They called it a short-term step while broader policy gets drafted [7].

But short-term steps have a way of becoming infrastructure. The precedents set this month, Anthropic's overnight shutdown and OpenAI's approval-gated launch, will shape how every future model release works. Once a government has a veto, it rarely gives it back.

What this means for builders

If you build on frontier model APIs, the rules changed this month. Here is the practical picture:

• Model availability is now a policy question, not a technical one. Your API access depends on geopolitical decisions, not uptime SLAs. Plan accordingly.
• Preview periods may be government-gated, not just company-gated. You cannot assume that early access means public access follows in two weeks. It might follow in two months, or with restrictions attached.
• Multi-model strategy is no longer optional. If your product only works with one provider's frontier model and that model gets pulled, you are down. Fallbacks, abstraction layers, and local models are insurance, not luxury.
• Open and local models are the hedge. A smaller model you control is worth more than a frontier model that requires government approval to use.

The irony at the center

Anthropic told Washington their model was a cyber weapon, and Washington believed them. A technology adviser to the White House, who co-leads the president's council on technology and science, said Anthropic's CEO "spiked the cortisol level" of officials by describing Mythos as a model with advanced cyber capabilities [8]. The government then built a gating framework around that fear.

OpenAI, watching this happen to its rival, chose to cooperate with the same framework rather than risk the same fate. That is rational. It is also how voluntary frameworks become mandatory ones without anyone passing a law.

A member of Congress and co-author of a bipartisan AI regulation bill put it plainly: "No law. No process. No oversight. Just appointees in Washington deciding who's in and who's out" [9].

Where we are now

Three weeks ago, frontier AI was a product you could buy. Today it is a controlled resource that the US government reviews before release. Two companies have now accepted this. No law required it. An executive order and some pressure did the job.

If you are building anything on these APIs, treat model availability as a supply chain risk. Diversify. Keep local fallbacks. And pay attention to policy news the same way you pay attention to changelogs, because this month they are the same thing.