News · 2026-06-27
The government cleared one Anthropic model and kept the other locked up
For two weeks, one of the most capable AI systems in the United States was simply switched off by order of the government, and this week it came partly back on. The U.S. Commerce Department partially cleared Anthropic's Mythos 5 model, allowing roughly a hundred American companies and federal agencies to use it again. Its more advanced sibling, Fable 5, remains fully blocked. If you want a single picture of where frontier AI sits in 2026, it is this: a private company built two of the smartest tools on the planet, and a government is now deciding, model by model and customer by customer, who is allowed to touch them.
To understand how unusual this is, you have to back up. For years, releasing an AI model meant putting it online for anyone to use or pay for. The competition was about capability - whose model was smartest, fastest, cheapest. Access was assumed. That assumption just broke. Earlier this month the government issued an export-control directive citing national-security authorities, and Anthropic responded by cutting off all access to both Mythos 5 and Fable 5 - including, remarkably, for its own employees who are foreign nationals. The Defense Department had already labeled Anthropic a "supply chain risk," a phrase normally reserved for foreign adversaries, not a California AI lab. Anthropic is suing the administration to reverse that designation, and it quietly moved co-founder Tom Brown into the lead negotiator role, stepping in front of chief executive Dario Amodei for the talks with Washington.
Here is a way to picture what changed. Imagine a power company that built a reactor capable of lighting up the whole region. In the old world, it sold electricity to anyone who plugged in. In the new world, a government inspector sits at the switchboard and decides which houses get connected, leaves the most powerful generator offline entirely, and labels the utility a national-security concern while the lawyers argue. The electricity didn't get worse. The question of who controls the wire became the whole story.
This didn't happen in isolation. The same week, OpenAI launched its new GPT-5.6 family but agreed to a government request to stagger the rollout, granting access to a small set of enterprise customers approved one at a time. Independent developer Simon Willison walked through the details of that gated launch. Two of the three leading American labs, both with their newest models throttled by the same administration in the same week - but treated very differently. OpenAI got a polite "request" and complied. Anthropic got an export directive, a supply-chain-risk label, and a lawsuit. That gap is not a footnote; it is the most strategically interesting thing on the board.
Why does this matter beyond the two companies involved? Because it changes what "the best AI" even means as a competitive advantage. If the most powerful models are available only to a government-approved list, then the moat is no longer just engineering talent or training compute - it is regulatory standing. Whoever the government trusts gets to ship. That is industrial policy, the kind of thing that usually governs jet engines and advanced chips, now applied to software you talk to. It also reshapes the global map: when American frontier models get harder to obtain, customers and competitors abroad have every reason to build or buy alternatives, which is exactly the spillover regulators say they want to avoid.
The community reaction has been loud and split. On forums devoted to long-term AI questions, the dominant read is alarm - states reaching to control advanced AI before it controls anything, with a libertarian "don't fence in the technology" camp on one side and a "some oversight is the lesser evil" camp on the other. The competitive-angle threads obsess over the Anthropic-versus-OpenAI disparity and what it signals about lobbying, compliance, and who has the better relationship with Washington.
The honest caveat: we are seeing this through press reports and a still-running lawsuit, not through any published rulebook. Nobody outside the negotiating room knows the actual criteria for why Mythos 5 cleared and Fable 5 didn't, or why two labs got different treatment. That opacity is itself the problem worth watching. A gating system with no public standard is hard to distinguish from favoritism, and the thing to track over the coming weeks is whether Fable 5 clears, whether Anthropic's suit forces any criteria into daylight, and whether "approved-customer list" quietly becomes a permanent feature of how frontier models ship. For more on why model weights have become this contested, see our explainer on open-weight models.