The Government Pulled the Plug on the World's Most Capable AI Models for 15 Days. Yesterday It Turned One Back On for 100 Companies.

On June 12 at 5:21pm ET, Anthropic received a legally binding US government export control directive and disabled Claude Fable 5 and Mythos 5 for every customer worldwide. The models had launched three days earlier. The trigger was a report from Amazon CEO Andy Jassy to Treasury Secretary Scott Bessent about a jailbreak technique that could unlock Mythos 5's advanced cybersecurity capabilities through Fable 5. Anthropic said it believed the concern was a misunderstanding. The directive was binding regardless.

The scope of the order is what makes it unprecedented. The government required Anthropic to suspend access not only for foreign nationals outside the US, but for foreign national employees working inside Anthropic's own offices. No carve-out for ongoing enterprise contracts. No grace period for customers running production systems. No distinction between API access and consumer products. Every deployment — medical, legal, financial, commercial — went dark simultaneously. This is the first time the US government has used export control authority to withdraw a commercially deployed AI model from a global market mid-deployment.

On June 26, the US Commerce Department granted Anthropic permission to restore Mythos 5 for roughly 100 approved companies and federal agencies. Commerce Secretary Howard Lutnick communicated the decision in a letter to Anthropic's chief compute officer. The letter grants access specifically for companies that operate and defend critical infrastructure. Fable 5 remains unavailable as of today. What the partial restoration reveals is the architecture of a new access framework: a tiered system in which the US government functions as the certifying authority for who can use frontier AI and under what conditions.

Enterprise software companies that built production systems on Fable 5 discovered, in real time, that their infrastructure has a dimension of risk with no analogue in the history of software. A database vendor can go bankrupt. A cloud provider can have outages. A regulatory shutdown that applies globally, instantly, and at government directive is a different category of counterparty risk entirely. The 15-day blackout will not be the last instance of this mechanism. The Mythos 5 framework — approved institutions, critical infrastructure focus, government-controlled access — is a template for how the US plans to manage frontier AI export risk going forward. AI companies building in verticals with national security adjacency should assume this framework will be applied to them.

The right frame for investors is not "will AI get regulated" but "in what currency will AI access be denominated." The Mythos 5 restoration creates a trusted-partner tier for frontier AI — a credential issued by a government department, not a vendor. That credential has value that scales with the capability of the underlying model. The companies with approved access to Mythos 5 now hold a competitive advantage that cannot be purchased, only granted. Whether this was the intended design of the policy or an accidental consequence matters less than the structural reality it creates: governments are becoming the gatekeepers of AI advantage, and the precedent was set this week.