The switch flipped back

On June 12, an export-control order pulled Anthropic's two frontier models offline for everyone on Earth. By July 1, one of them was back. The Commerce Department re-authorized Mythos 5 on June 26-27 for a defined set of roughly 200 US critical-infrastructure defenders — Apple, Google, Cisco, Nvidia, Microsoft among them — and lifted Fable 5's restrictions on June 30, returning it globally the next day across Claude.ai, the API, Claude Code, and Cowork.

The retrained classifier blocks the reported jailbreak in over 99% of attempts, quietly rerouting anything it flags to the older Opus 4.8. — Fortune

What unlocked it was a concession, not a court ruling. Anthropic retrained the safety classifier that Amazon had reported jailbroken, and says the new version blocks that specific attack in over 99% of attempts, quietly rerouting anything it flags to the older Opus 4.8. Commerce's own testers — its Center for AI Standards and Innovation — checked the old and new safeguards and called them extraordinarily strong before the models were allowed back.

This is the other half of an arc no frontier model had traced before: a government switched a live commercial model off, then switched it back on, on terms it dictated. The template now exists. It also carries a cost the vindication headlines skip — the same testers noted the tougher classifier throws more false alarms on ordinary coding and debugging, so the security guarantee buys a model that now sometimes refuses legitimate work. And while the models sat dark, the market moved on without them: OpenAI's newest cyber model overtook Mythos 5 on a standard vulnerability benchmark, and Anthropic could not answer because its model was, by federal order, undeployable.