Claude Fable 5 Suspended on US Export Order

Anthropic has disabled its most advanced AI model just three days after its release. Citing national security concerns, the US government issued an export control directive ordering the company to restrict access to Claude Fable 5 and Mythos 5. In response, Anthropic disabled both models for all customers on June 12, while leaving other Claude models unaffected.

The directive specifically targeted foreign nationals, barring them from using Claude Fable 5 and Mythos 5 regardless of whether they were inside or outside the US—a restriction that also applied to Anthropic’s own foreign employees. Because blocking access selectively by nationality proved technically unfeasible, the company chose to withdraw the models entirely.

The sudden suspension comes amid notable irony. Only days earlier, users had complained that the model's safeguards were too aggressive, frequently routing harmless prompts to older models. Now, security vulnerabilities within those same safeguards have served as the catalyst for the shutdown.

Jailbreak Concerns Trigger Shutdown

The immediate catalyst for the order was concerns over potential "jailbreaking" techniques. The export control directive arrived at 5:21 p.m. ET on June 12 without specifying the underlying security rationale. In an official statement, Anthropic noted it understands the government became aware of a specific method to bypass Claude Fable 5’s safeguards.

However, the company’s internal review of the vulnerability painted a different picture. The issues identified consisted of a few minor, pre-existing vulnerabilities that other public models, including OpenAI’s GPT-5.5, would flag just as easily without requiring a jailbreak. According to Anthropic, the government’s evidence consisted only of a verbal description of a narrow attack vector: prompting the model to analyze a specific codebase and resolve its defects. The company emphasized that over 1,000 hours of pre-launch bug bounty programs and external red-teaming failed to surface any universal jailbreak.

Anthropic has strongly contested the decision, arguing that suspending a commercial model used by hundreds of millions over a narrow, hypothetical vulnerability is unwarranted. The company warned that enforcing such a standard would effectively halt the deployment of any new frontier AI models. While Anthropic acknowledges the government's export control authority to block unsafe releases, it maintains that such actions must be executed through transparent, fair, and technically sound legal channels.

Further complicating the situation is the origin of Mythos 5. The model was initially deployed for cybersecurity defense under Project Glasswing, a collaborative initiative with the US government. Having co-developed the model, the government has now ordered its withdrawal.

Microsoft's Ban and the 30-Day Data Retention Policy

Even before the federal directive, Claude Fable 5 faced adoption hurdles. Microsoft had already barred its own employees from using the model, removing it from the internal GitHub Copilot selection menu. The restriction was to remain in place pending a legal and security review of the model's data retention policies.

Paradoxically, Microsoft continued to offer the same model to its corporate clients through Copilot and its Foundry platform. While external customers were permitted to use the model, internal employees remained restricted until the company could verify the destination of the data—highlighting the divide between commercially available AI and internally approved software.

The primary point of contention is Anthropic's 30-day data retention policy. Both Fable 5 and Mythos 5 require all user traffic to be stored for 30 days, offering no exemptions even for enterprise clients with pre-existing zero data retention (ZDR) agreements. Anthropic maintains that this data is used solely for jailbreak detection and reducing false positives, rather than model training. However, the data retention policy lets reviewers read conversations when severe harm is suspected, or upon written client request, raising concern among corporate security teams.

The Uncertain Future of Claude Fable 5

Regardless of corporate caution over data retention, Fable 5's initial release was designed to be temporary. Anthropic had announced that the model would be provided at no additional cost to Pro, Max, Team, and seat-based Enterprise subscribers only until June 22. On June 23, the model was scheduled to be transitioned to usage-based pricing, with plans to reintegrate it into standard subscriptions once computing capacity allowed.

The usage-based rates were set at $10 per million input tokens and $50 per million output tokens—representing twice the cost of the older Opus 4.8 model, but less than half the rate of Mythos Preview. However, the export control order has blocked this planned commercial pathway entirely.

While Anthropic has characterized the shutdown as a misunderstanding and promised a detailed explanation within 24 hours—along with a pledge to restore access as quickly as possible—the precedent has been set. This marks the first time a government has forced the withdrawal of Claude Fable 5, a frontier AI model, just three days post-launch. The long-term implications for the industry now depend on export control decisions in Washington, rather than the developer's roadmaps.