Anthropic's strongest models are back, but not back in the old sense. They now return through a government approval workflow.

The paper's May feature said Anthropic's moral voice had met its compute bill, while Wednesday's OpenAI story showed frontier cyber capability moving behind verified-defender access. Anthropic's Fable and Mythos redeployment joins those tracks. The model is not only a product. It is a permissioned surface.

Anthropic says the U.S. government applied export controls to Claude Fable 5 and Claude Mythos 5 on June 12, requiring the company to restrict access to foreign nationals inside and outside the United States. Because it had no reliable way to verify nationality in real time, Anthropic suspended access to both models for all users. [1]

That operational detail is the quiet bomb in the announcement. A nationality rule became a universal outage because the identity layer was not ready for the legal distinction. The model did not disappear only from prohibited users. It disappeared from everyone while the company and the government figured out how the permission boundary could be enforced. [1]

On June 30, Anthropic said the export controls had been lifted. Fable 5 would become available globally starting July 1 on the Claude Platform, Claude.ai, Claude Code, and Claude Cowork, with plan-specific weekly usage limits through July 7 and access on AWS, Google Cloud, and Microsoft Foundry to be re-enabled as quickly as possible. Mythos 5, by contrast, was restored only for a set of U.S. organizations after government approval on June 26, while Anthropic continued coordinating with the government to expand access through the Glasswing program. [1]

That split is the whole story. Fable returns globally. Mythos returns to approved organizations. The commercial product and the security product now run through different gates.

CNBC framed the announcement as the Trump administration lifting export controls on Claude Fable 5 and Mythos 5, ending the latest standoff between Anthropic and the government. It noted criticism from tech executives and investors who argued the crackdown handed time to Chinese open-source developers trying to catch up. [2] That is the market view: a restriction lifted, a competitive clock resumed.

The institutional view is more important. Anthropic's own post says the June 12 directive followed an Amazon researcher report describing a way to bypass Fable 5 safeguards so the model identified vulnerabilities and, in one case, produced exploit-demonstration code. Anthropic says testing found many less capable models could identify the same vulnerabilities, and all models it tested could produce the same demonstration. It says the technique did not expose unique Mythos-level capabilities, but it trained a new classifier that blocks the behavior in more than 99 percent of cases. [1]

X sees the redeployment as either censorship retreat, capture, or proof that AI labs have become regulated utilities without the public admitting it. Mainstream coverage says export controls were lifted. The gap is the approval workflow itself. Model access now depends on the Commerce Department, partner reports, safety classifiers, Glasswing organizations, cloud platforms, and a lab's ability to sort users by national status and role.

That does not make the workflow illegitimate. It may be the only practical way to handle a model that Anthropic describes as unusually strong for cybersecurity. But it does make the workflow political. A model gate decides who can defend critical systems, who must wait, whose country matters, and which partner's report can freeze a global product.

The useful comparison is not to software updates. It is to controlled infrastructure. A lab ships a capability; the state intervenes; partners test it; a classifier changes; approved users return first; cloud channels follow. That is a commerce-control regime, not a launch note.

Anthropic's moral language has always asked the public to trust its restraint. The new receipt asks a harder question: who supervises restraint when the model, the export-control office, the cloud vendor, and the approved customer all sit inside the same operational chain? Fable is back. Mythos is allowed. Those are not the same verb.

-- THEO KAPLAN, San Francisco