The U.S. government is set to take an awful lot of control over which AI models get released.

Two weeks after the U.S. government pulled Anthropic’s Fable and Mythos models, OpenAI’s new model seems to be headed for the same limbo. The Information broke the news Thursday that GPT 5.6 would be released only into limited preview, with the government approving the release “customer by customer” until a general release can be approved.

If that preview only lasts a “couple of weeks,” as Altman reportedly projected, that might not be a particularly big problem. But Mythos has already been in preview for months, and there’s no indication it will make it to general release any time soon. Even a few weeks spent in review could significantly limit the economic upside of a costly new system, at a time when AI labs are trying desperately to improve their bottom lines. If the pace of model development slows as a result, it’s likely to put a similar chill on the ongoing data center buildout.

If this goes bad, the entire industry could be at risk.

Critically, OpenAI and Anthropic are now in the same exact position with the same problems facing them and the same disaster waiting if they fail. Conversations within the tech industry tend to focus on the role of one side or another in bringing this on, either accusing Anthropic of running a regulatory capture scheme or accusing OpenAI of cozying up to Trump to ice out a rival. It’s understandable; many of the most prominent people in the industry have billions of dollars riding on one company or the other. 

But what’s happening now is bigger than that. The cost of implementing a haphazard government approval process for every frontier model is obvious, and there’s no fix that helps one lab without helping the others. 

The most immediate problem is simply establishing a release process that makes sense. It’s fine for the government to test models before release (this is how it works for lots of consumer products)— but as GMU fellow (and soon-to-be OpenAI employee) Dean Ball detailed in an eloquent post this morning, it’s not clear what kind of safety assurances could be put in place to satisfy regulators. The U.S. government doesn’t have the expertise or capacity for the kind of testing that would be needed here. It’s not even clear what regulators would be trying to protect against, since there’s been no effort to articulate what risks the government is actually concerned about.

It’s tempting to see the government process as the whole of the problem itself, but there are real concerns underneath. Even if you don’t believe the Mythos hype, there’s clear evidence of how AI tools are revolutionizing cybersecurity. There are similar processes at work in biorisk and alignment. Restricting model releases can’t be the whole answer in itself — that will only limit what’s available to the public — but there are real concerns to be addressed.

The best ideas for addressing them, as laid out by Ball, will mean working together. It will mean trusting independent groups to guide the process, even if they don’t completely align with your goals. It will mean lining up behind the least-bad regulatory options available, instead of fighting every regulation tooth-and-nail. And most of all, it will mean fighting for AI as an industry, instead of seeing safety and regulation as opportunities to gain an advantage.

For a lot of people working in AI, that will be a tough sell. Unfortunately, AI models have progressed to the point where their capabilities have real political consequences. Dealing with those consequences will require collective action. In the weeks to come, we’ll find out if that’s something the industry is capable of.

View Bio