The US government is poised to heavily control which AI models will be released.
Two weeks after the US government withdrew Anthropic’s Fable and Mythos models, OpenAI’s new model seems headed for the same void. The Information made the news on Thursday that GPT 5.6 will only be released in limited preview, with the government approving a “customer-by-customer” release until a general release is approved.
If this preview only lasts “a few weeks,” as Altman reportedly suggested, that might not be too much of a problem. But Mythos has already been in preview for months, and there’s no sign that it’ll hit general release anytime soon. Even a few weeks spent on revision could significantly limit the financial upside of an expensive new system as AI labs desperately try to improve their results. If the pace of model development slows as a result, it is likely to bring a similar chill to continued data center growth.
If this goes badly, the entire industry can be at risk.
Crucially, OpenAI and Anthropic are now in exactly the same position with the same problems they face and the same disaster that awaits if they fail. Debates within the tech industry tend to focus on one side or the other’s role in making this happen, either accusing Anthropic of running a regulatory capture system or accusing OpenAI of aspiring to Trump’s freezing of an opponent. It’s understandable. Many of the most prominent people in the industry have billions of dollars with one company or another.
But what is happening now is bigger than that. The cost of implementing a haphazard government approval process for each border model is obvious, and there is no solution that helps one lab without helping the others.
The more immediate problem is simply creating a release process that makes sense. It’s fine for the government to test models before release (that’s how it works for many consumer products)—but as GMU contributor (and future OpenAI employee) Dean Ball explained in an eloquent post this morningit is not clear what kind of safety guarantees could be put in place to satisfy regulators. The US government does not have the expertise or capacity for the kind of testing that would be needed here. It’s not even clear what the regulators would be trying to protect against, as no attempt has been made to articulate the risks the government is actually worried about.
It’s tempting to see the government process as the whole of the problem, but there are real concerns underneath. Even if you don’t believe the Mythos hype, there is clear evidence of how AI tools are revolutionizing cybersecurity. There are similar processes in place biorisk and alignment. Limiting model releases can’t be the whole answer by itself – that will only limit what’s available to the public – but there are real concerns that need to be addressed.
The best ideas for dealing with them, as articulated by Ball, will mean collaboration. It will mean trusting independent teams to guide the process, even if they don’t fully align with your goals. It will mean lining up behind the least bad regulatory options available, rather than fighting every regulation tooth and nail. And most of all, it will mean fighting for AI as an industry, rather than seeing security and regulation as opportunities to gain an advantage.
For many people working in AI, this will be a tough sell. Unfortunately, AI models have advanced to the point where their capabilities have real policy implications. Addressing these consequences will require collective action. In the coming weeks, we’ll find out if this is something the industry can do.
When you purchase through links in our articles, we may earn a small commission. This does not affect our editorial independence.
