White House Now Gates Frontier AI Access at OpenAI, Anthropic

The interesting shift, playing out over the last few weeks, is that the Trump administration has quietly become the gatekeeper on the customer list at the two leading U.S. frontier labs. According to reporting compiled by The Information and matched by CNBC, OpenAI's newest product, called GPT-5.6 Sol, is being made accessible only to customers approved by the Trump administration. Anthropic's Fable 5 and Mythos 5 were briefly taken offline to comply with a directive blocking their use by foreign nationals, with Mythos 5 later redeployed to what was described as a small group of cyber defenders and infrastructure providers.

This is a real departure from what the administration was selling only weeks earlier, when the message to the industry was that regulation would not be allowed to impede innovation. Dean Ball, a co-author of Trump's AI Action Plan and the incoming leader of OpenAI's new Strategic Futures team, put it more bluntly than anyone in government has, saying U.S. federal AI policy has gone 'from implausibly libertarian to increasingly draconian and opaque' in a matter of weeks. Google DeepMind's Demis Hassabis has meanwhile been pitching a FINRA-style industry body to formalize whatever this is turning into, while Anthropic's Dario Amodei has floated an FAA-style agency with direct authority to block unsafe models.

Why this matters if you are a buyer rather than a builder: the model on your roadmap now has a second approval gate that has nothing to do with your contract with the vendor. Foreign-national employees appear to be a trigger. Cybersecurity and critical-infrastructure use cases seem to move to the front of the queue. Everything else sits in a queue no one has published rules for.

The honest caveat is that this is a synthesis piece reading across a series of prior scoops rather than announcing a single new decision, and the administration's public line is still that there will not be an AI equivalent of the FDA, as former White House AI policy adviser Sriram Krishnan told the Financial Times on July 5. What the reporting does not give you is the actual approval criteria, any appeals process, or whether Google, xAI and Meta are being handled the same way. If you are running procurement right now, the safer planning assumption is that frontier access timing is a political variable, and it is worth carrying a non-frontier vendor in the stack.