The United States government appears to be taking growing control over which advanced artificial intelligence (AI) models are allowed to reach the market. Two weeks after the U.S. government pulled Anthropic’s Fable and Mythos models, OpenAI’s newest model now faces similar uncertainty.
What happened
The Information reported on Thursday that GPT‑5.6 will be released only as a limited preview, with the government approving access “customer by customer” until a broader release is permitted. If that preview lasts only a “couple of weeks,” as Sam Altman reportedly suggested, it may not be a major issue. However, Anthropic’s Mythos has already been in preview for months, and there is no sign it will be approved for general release soon.
Why this matters economically
Even a few weeks of review can materially reduce the economic upside of an expensive new system, at a time when AI labs are struggling to improve margins. If model development slows as a result, it is likely to chill ongoing data center buildouts.
The piece warns that a failed approval process could threaten the whole industry. OpenAI and Anthropic now find themselves in essentially the same position, facing identical problems and similar downside risk.
The debates and competing interests
Industry conversations often focus on who is to blame for the situation — some accuse Anthropic of attempting regulatory capture, while others accuse OpenAI of cultivating political ties to disadvantage a rival. That tension is understandable: many prominent industry actors have billions of dollars tied to the success of one company or another.
The core problem
At root, the most immediate issue is the absence of a sensible release process. It is reasonable for the government to test models before release — many consumer products undergo premarket review — but it is unclear what kinds of safety assurances would satisfy regulators for cutting‑edge AI models.
Dean Ball, a George Mason University (GMU) fellow and a soon‑to‑be OpenAI employee, argued in a detailed post this morning that the U.S. government lacks the necessary expertise and capacity to perform the kind of testing that would be required. It is also unclear what specific risks regulators are trying to guard against, since those risks have not been clearly articulated.
What risks are at stake?
Even if one discounts the hype around Mythos, there is clear evidence that AI tools are transforming cybersecurity. Similar dynamics exist around biorisk and alignment. Restricting model releases cannot be the only answer—doing so would simply limit what is available to the public—yet the underlying concerns are real and need addressing.
Proposed approaches and the need for collective action
According to Ball, the best path forward involves collaboration: empowering independent groups to help guide the process even if they do not fully align with every stakeholder’s goals. It also means accepting the least‑bad regulatory options rather than opposing every regulation outright. Above all, the industry must act collectively for the health of AI as a sector rather than treating safety and regulation purely as opportunities to gain advantage.
For many working in AI, that will be a difficult proposition. The capabilities of modern models now carry real political consequences. Handling those consequences will require coordinated action. In the coming weeks, it will become clearer whether the industry can muster that cooperation.
Summary
The U.S. government’s emerging model‑approval practice in its current form could limit which AI systems reach consumers and how quickly. That has economic implications for AI firms and data center investments and highlights the need for cross‑industry, independent solutions to manage genuine risks.
