The US government will have tremendous control over which AI models it releases.
Two weeks after the U.S. government retired Anthropic’s Fable and Mythos models, OpenAI’s new model appears to be heading toward the same dead end. The Information broke the news on Thursday that GPT 5.6 will only be released as a limited preview, with the government approving releases on a “customer by customer” basis until it is approved for general release.
Even if the preview only lasts “a few weeks,” as Altman predicts, that might not be a big deal. However, Mythos has already been in preview for several months and there is no sign of it being released to the public anytime soon. At a time when AI labs are desperately trying to improve their bottom line, even a few weeks of review can severely limit the economic benefit of an expensive new system. As a result, a slowdown in model development could have a similar negative impact on ongoing data center construction.
If this worsens, the entire industry could be at risk.
Importantly, OpenAI and Anthropic are now in exactly the same position, facing the same problems and facing the same disaster if they fail. Conversations within the tech industry tend to focus on the role of either side in causing this problem, either accusing Anthropic of carrying out a regulatory capture scheme or accusing Open AI of coddling Trump to eliminate its rivals. That’s understandable. Many of the most prominent people in the industry have billions of dollars in fortunes in one company or another.
But what’s happening now is bigger than that. The cost of implementing a haphazard government approval process for all frontier models is clear, and there is no solution that helps one lab without helping the others.
The most pressing issue is simply establishing a meaningful release process. It’s fine for governments to test models before release (this is how many consumer products work), but as GMU fellow (and soon-to-be OpenAI employee) Dean Ball detailed in an eloquent post this morning, it’s not clear what safety guarantees can be put in place to satisfy regulators. The U.S. government does not have the expertise or capacity for the type of testing needed here. It is not even clear what regulators are trying to protect against, as there has been no effort to clarify what risks the government is actually concerned about.
It’s tempting to see the government process as a whole, but there are real concerns at the root of it all. Even if you don’t believe the Mythos hype, there is clear evidence of how AI tools are revolutionizing cybersecurity. A similar process is at work in biorisk and coordination. Limiting the release of models is not the answer to everything in itself, it just limits what is made available to the public. But there are real concerns that need to be addressed.
The best idea for dealing with them, as Ball presented, means working together. It means trusting an independent group to guide you through the process, even if it doesn’t align perfectly with your goals. It means lining up behind the least harmful regulatory option available, rather than fighting every regulation outright. And above all, it means fighting for AI as an industry, rather than seeing safety and regulation as an opportunity to gain an advantage.
For many people involved in AI, this will be difficult. Unfortunately, AI models have advanced to the point where their capabilities have real-world political consequences. Addressing those consequences requires collective action. The coming weeks will tell whether that is possible for the industry.
If you buy through links in our articles, we may earn a small commission. This does not affect editorial independence.
