Grounded

Claude Fable 5 lasted three days before the Commerce Department pulled it from service worldwide.

Anthropic released Claude Fable 5 last Monday, the most capable AI model it had ever made available to the public. By Thursday evening, the US government had ordered it pulled from service worldwide. The intervening 72 hours produced a developer backlash over hidden capability restrictions, a public apology, and an export control directive from the Commerce Department citing an unspecified national security concern, a sequence without precedent in the AI industry.

The invisible wall

Anthropic launched Claude Fable 5 on June 9 as the first publicly available model in its new Mythos class, a tier the company created above Opus for its most capable systems. Mythos had existed since April in limited release through Project Glasswing, restricted to organizations managing critical infrastructure; Fable 5 wrapped that same architecture in safety classifiers and opened it to API customers and subscribers. Those classifiers routed requests involving cybersecurity, biology, chemistry, and model distillation to Claude Opus 4.8, and told the user when that happened.

One category of safeguard worked differently. A paragraph on page 13 of Fable 5’s 319-page system card disclosed that when the model detected requests it classified as frontier LLM development work, including efforts to distill its capabilities into smaller models, it would degrade its own output without notification. The user received a weaker answer with no indication that the full model had been withheld.

Developers found the hidden behavior within hours. Nathan Lambert of the Allen Institute for AI posted that having his access to cutting-edge models “rug pulled in an under the table fashion” was “appalling.” Fortune reported accusations of “secret sabotage,” and The Register documented the model refusing innocuous prompts outright. Anthropic apologized within two days, called the invisible safeguard “the wrong tradeoff,” and committed to making all fallbacks visible, with explicit notification when a request was being handled by Opus 4.8 instead of Fable 5. The underlying restriction remained; the change was transparency, not removal.

Immediate shutdown

That backlash was still fresh on June 12 when the Commerce Department intervened. At 5:21 PM Eastern, an export control directive arrived from Secretary Howard Lutnick ordering Anthropic to suspend all Fable 5 and Mythos 5 access for any foreign national, whether inside or outside the United States. The government’s stated concern was a jailbreak it believed could bypass Fable 5’s safety classifiers, though the directive provided no specific details of the national security risk involved.

The restriction extended to Anthropic’s own foreign national employees, making ordinary service operationally impossible. Rather than attempt selective enforcement it could not guarantee, Anthropic disabled both models for all users worldwide. All other Claude models, including Opus 4.8, Sonnet 4.6, and Haiku 4.5, remained fully operational; by 9:59 PM Eastern, API calls to Fable 5 and Mythos 5 returned 404 errors directing users to Opus 4.8.