Anthropic Got the Green Light. OpenAI Got a Leash. Same Government.

Two AI companies. One administration. Two very different outcomes.

Anthropics's Mythos 5 has been cleared for deployment across more than 100 US companies and government agencies — including, notably, their non-American employees. OpenAI, meanwhile, launched its GPT-5.6 series in limited preview and simultaneously acknowledged that the government had asked it to restrict the rollout. The models exist. The capability is there. But one company is expanding access and the other is publicly complaining about being told to hold back.

OpenAI didn't stay quiet about it, either. The company said, in plain terms, that it doesn't believe this kind of government access process should become the long-term default — that it keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them. That's a pointed statement. That's a company describing a constraint it considers unjust while being forced to live inside it.

Two Models, Two Answers

Here's what makes this strange: both companies are making serious safety claims. OpenAI describes GPT-5.6 Sol as having its most robust safety stack to date, with protections for high-risk activity, sensitive cyber requests, and misuse — tested for weaknesses, hardened against real-world attacks. Anthropic, per reporting from 9to5Mac, had actually been forced to disable access to both Claude Fable 5 and Claude Mythos 5 for all customers following an export control directive two weeks earlier. Then came a deal. Then came clearance. A partial victory, as one outlet framed it, negotiated rather than granted.

So the sequence, compressed: Anthropic gets shut down, negotiates back in. OpenAI launches, gets asked to limit access, says publicly that this shouldn't be how things work. One company got a yes. The other got a conditional no and a leash long enough to complain about.

The models themselves are different products — Mythos is positioned around cybersecurity, while GPT-5.6 Sol touts agentic improvements in coding, biology, and cybersecurity both. But the regulatory treatment doesn't map cleanly onto capability differences or safety profiles. It maps onto outcomes that are hard to explain by policy logic alone.

The Consistency Problem

National security policy, when it's functioning, looks boring — same rules, applied the same way, to similar risks. What we're watching here doesn't look boring. It looks like two companies operating in the same space, under the same government, getting sorted into different buckets by a process that hasn't explained its own criteria.

That matters beyond the companies involved. If the framework for what AI gets cleared and what gets restricted can't be read from the outside — if Anthropic gets restored access after a negotiation and OpenAI gets told to wait with no clear rationale for the asymmetry — then the policy isn't really a policy. It's a series of individual decisions dressed up in the language of national security.

OpenAI's public statement reads less like a complaint and more like a legal brief being filed in the court of public opinion. They're not just frustrated. They're building a record. And Anthropic, for its part, isn't gloating — getting access stripped and then partially restored two weeks later is not a clean win. It's a reminder of how quickly the ground can shift.

The administration has demonstrated that it has opinions about which AI models reach which hands. What it hasn't demonstrated is a consistent theory of why.

That gap is where trust goes to die.