The United States has quietly imposed a de facto licensing requirement on the most advanced artificial intelligence systems, compelling leading developers to delay or restrict access to their latest models. Under a January 2026 executive order, OpenAI’s GPT-5.6—comprising the Sol, Terra, and Luna variants—is now available only to a handful of pre-approved partners, while Anthropic’s Fable 5 was withdrawn entirely after the government demanded it block access by foreign nationals. Both companies confirmed they are complying despite public unease; OpenAI stated it does “not believe that this government access process should become a long-term setting.”

Viewed from Washington, the move formalises an oversight mechanism that the AI industry had previously navigated through voluntary agreements. Former White House AI adviser Dean Ball described it as a compulsory licensing regime that creates “excessive restrictions,” warning that the absence of clear safety standards could delay launches and advantage China’s AI efforts. The controls arrive at a moment when the commercial AI sector is already grappling with the economic and operational consequences of its own rapid expansion.

Across European and Asian technology hubs, companies are confronting a stark reversal from the earlier “use as much as you can” ethos. Consultancy Accenture has begun preventing employees from burning token quotas on trivial tasks, after discovering that agentic AI loops—where systems recursively self-check and retry—cause unpredictable and exponential cost surges. Meanwhile, purpose-built healthcare AI systems from OpenAI and Anthropic, which claim to rival human physicians on certain benchmarks, remain hamstrung by the absence of independent validation; OpenAI’s own internal testing, conducted with 260 physicians across 60 countries, reduced medical error rates by 71 percent, but critics argue that simulated environments do not replicate real clinical complexity.

Longitudinal studies reinforce the stakes. MIT researchers found that programmers using AI tools completed work faster but were weaker at diagnosing novel errors; emergency physicians at a University of California study showed slower clinical decision-making when automated diagnostic systems were removed; and Stanford documented students’ increased willingness to accept plausible-sounding but incorrect AI answers. These findings underscore a cognitive trade-off that policymakers are only beginning to factor into regulatory design.

The immediate milestone to watch is the duration and scope of the US government’s review framework. OpenAI has indicated the restrictions are temporary, but no sunset clause has been published. As the industry prepares for a potential wave of similar regulations, the balance between innovation, cost discipline, and cognitive integrity will shape the next phase of AI deployment.