US Gov Now Gates AI Releases: GPT-5.6 in Limbo

The industry's internal conversation has mostly been about blame. One camp accuses Anthropic of running a regulatory-capture play. Another accuses OpenAI of cozying up to the administration to freeze out a competitor. Both narratives are understandable, given how many prominent people have billions riding on one company or the other. But the structural problem is bigger than any single rivalry, and pointing fingers does nothing to resolve it.

How a customer-by-customer approval actually works

The mechanics here are unusual. Instead of a model launching to the general public on a set date, access is granted in pieces. The government signs off on individual customers being allowed to use the model, one approval at a time, until a broader general release is authorized.

If that preview phase is short, the damage is limited. Altman reportedly projected GPT-5.6's preview might last only a couple of weeks. The trouble is that Mythos is the cautionary example: months in, still no general release.

Even a few weeks of review can erode the economic case for an expensive new system. AI labs are pouring enormous sums into training frontier models and are under pressure to improve their bottom lines. A model that sits in regulatory limbo is a model that is not earning back its training cost, at precisely the moment the labs need it to.

If the pace of model development slows because releases keep stalling in review, the chill is unlikely to stop at the labs. It could ripple straight into the ongoing data center buildout that the entire AI economy is currently betting on.

The deeper problem: nobody has defined the rules

It is reasonable for a government to test products before they reach the market. That is how plenty of consumer goods already work. The difficulty is that it is not clear what kind of safety assurances would actually satisfy regulators here.

By the account of GMU fellow Dean Ball, who is reportedly soon to join OpenAI, the U.S. government does not currently have the expertise or the capacity to run the kind of testing that frontier model review would require. Worse, there has been no clear articulation of what specific risks the process is meant to guard against. Regulators are being asked to approve or block models without a stated definition of what counts as unsafe.

That said, treating the approval process as the entire problem misses something. There are genuine concerns underneath it. There is real evidence of AI tools reshaping cybersecurity, and similar dynamics are at work in biorisk and in alignment research. Simply restricting releases cannot be the whole answer, because that only limits what the public can access without addressing the underlying risk. But the risks are not imaginary.

What happens over the next 24 to 72 hours

The immediate question is GPT-5.6's preview timeline. Watch for any signal on whether the customer-by-customer phase stays measured in weeks or starts to resemble the open-ended limbo Mythos is in. That single data point will tell the industry a great deal about how this regime is going to operate.

• Expect continued reporting on which customers receive early GPT-5.6 access, since the approval list itself is now a meaningful signal.
• Watch for public statements from OpenAI and Anthropic. Whether they frame this as a shared industry challenge or revert to blaming each other will indicate if collective action is realistic.
• Look for commentary from policy figures like Ball on what a workable release process could look like, including the role of independent testing groups.
• Track whether any concrete framework for what regulators are testing against begins to emerge, because right now that definition is the missing piece.

The path forward, by Ball's framing, runs through cooperation that the industry is not used to. It would mean trusting independent groups to help guide the process even when they do not fully align with a given lab's goals. It would mean lining up behind the least-bad regulatory options rather than fighting every rule. And it would mean defending AI as an industry instead of treating safety and regulation as levers for competitive advantage.

For a lot of people working in AI, that is a hard sell. But the capabilities have advanced far enough that they now carry real political consequences, and managing those consequences will require collective action. The coming weeks will show whether the industry is capable of it.

Source: TechCrunch