The latest AI access fights are a warning about process, not just technology. Anthropic's Fable 5 return and OpenAI's GPT-5.6 preview show that the government can now shape who receives frontier models and when. That power may be necessary, but it is not yet disciplined enough to be trusted as infrastructure policy.
The Verge reported both Anthropic's restoration plan and OpenAI's staged release under federal pressure. Axios described the Anthropic episode as a 20-day standoff involving senior officials and uncertainty over global access. The lesson is not that the government should never intervene. A frontier model can plausibly affect cyber operations, weapons design, intelligence work, and strategic competition. Pretending otherwise is unserious.
The problem is improvisation. If access turns on emergency calls, unclear export directives, or case-by-case approvals that customers cannot understand, the United States risks weakening the ecosystem it wants to lead. Enterprises will hesitate to build on models that can vanish. Allies will seek sovereign substitutes. Smaller developers will be disadvantaged if only the largest labs can negotiate exceptions.
A workable system would publish model-risk categories, require independent evaluations, define trusted deployment environments, and explain when nationality, geography, sector, or use case matters. Some details can remain classified, but the governing structure cannot be a black box. The next frontier launch should test a rulebook, not the personal networks of lab executives.