Anthropic's frontier AI model Mythos 5 was partially cleared Friday for roughly 100 critical infrastructure operators; the same day, OpenAI delayed GPT 5.6 under the same regime. Two labs, one gate, and the public still has no idea how the selection is made.
The U.S. government's grip on frontier AI just tightened in a way that has less to do with any single model and more to do with who decides who gets one. On Friday, the Trump administration gave Anthropic permission to redeploy its most powerful model, Mythos 5, to a curated list of critical-infrastructure operators. The same day, it asked OpenAI to hold GPT-5.6 back from public release and ship it only to authorized partners whose identities were shared with authorities. Two labs, two products, one approval regime.
The pattern is the story. For roughly two weeks, Anthropic's Mythos 5 had been suspended over national-security concerns about military and intelligence misuse by adversaries including China and Russia. The partial reinstatement, reported by Reuters and confirmed by Wired and Politico, covers about 100 U.S. entities described as trusted organizations, including several Fortune 500 companies, according to an anonymous source cited in the Reuters wire. OpenAI's GPT-5.6 was delayed on the same day under a separate U.S. government request and will reach only partners whose names were passed to officials.
Anthropic described the partially restored model as "our most robust cybersecurity model" in its statement, and said it is working with the government to widen access. OpenAI has not publicly characterized its own model or the reasoning behind the delay. The parallel restrictions suggest Washington is treating dual-use AI risk as a coordinated policy frame, not a vendor-specific problem. That distinction matters. A one-off national-security intervention against a single company can be litigated, negotiated, or waited out. A regime that touches the two most prominent American labs simultaneously is harder to dismiss as an exception.
The new gatekeeping has no published criteria, no comment period, and no appeal process. John Coleman, legislative counsel at the Foundation for Individual Rights and Expression, called that out plainly: "No one knows how these companies are picked and why everyone else is excluded," he said, warning that concentrating that much power without explanation threatens the rule of law. OpenAI CEO Sam Altman posted similar concerns on X, agreeing that "extensive safety testing is not a bad idea" while objecting to government influence over who gets access in the first place.
What to watch next is whether the regime stays informal. If the same back-channel approvals extend to a third major lab, or if any rejected applicant publicly contests its exclusion, the current arrangement will start to look less like emergency export control and more like a standing licensing regime for the most capable AI. That is a category that, until this week, did not formally exist in U.S. policy.