Anthropic restores Claude Fable 5 after US export restrictions are lifted

On July 1, 2026, Anthropic restored access to Claude Fable 5 after US export restrictions were lifted. At the same time, it introduced a new cybersecurity classifier that blocks the jailbreak technique described in Amazon's report in more than 99% of cases.

How the new classifier works

The classifier screens incoming requests to Claude Fable 5. When it detects signs of an attack, it does not simply reject the request but reroutes it to Opus 4.8. This lets Anthropic keep the powerful Fable 5 available to legitimate users while reducing the risk that a known bypass technique is successfully exploited. According to the company, blocking effectiveness exceeds 99%: the vast majority of attempts to use the technique are filtered out at the entry point, before the request reaches the model.

• Access to Claude Fable 5 restored on July 1, 2026
• The earlier pause was caused by US export restrictions, which have been lifted
• The new classifier blocks the jailbreak technique from Amazon's report in more than 99% of cases
• Suspicious requests are automatically switched to Opus 4.8
• Anthropic proposed to Amazon, Microsoft and Google a joint four-criterion system for rating jailbreak severity

Why export restrictions affected access to the model

US export controls on frontier AI models have remained, in recent years, one of the main tools Washington uses to regulate the spread of dual-use technology. For companies like Anthropic, this means access to the most powerful model versions can be restricted temporarily by regulatory decision, not only for technical or commercial reasons. Fable 5's return on July 1, 2026 shows that such restrictions are not necessarily permanent: by lifting them, the regulator effectively let the company offer the model in full again.

Why a shared jailbreak rating system matters

Anthropic, together with Amazon, Microsoft and Google, has launched an initiative to standardize the rating of jailbreak severity across four criteria. Each company currently uses its own internal scales for how dangerous a given technique for bypassing model safeguards is, which makes it harder to compare risks across systems and to coordinate an industry response to new attacks. A single four-criterion system should give the industry a common language for describing the severity of incidents and could, possibly, form the basis for future industry safety standards.

Until now, leading AI labs have published data on their models' resistance to jailbreaks separately and in different formats, which made it impossible to compare their defenses directly. The joint initiative by Anthropic, Amazon, Microsoft and Google is meant to close precisely that gap.

What this means

The return of Claude Fable 5 after the lifting of export restrictions shows that regulatory barriers to Anthropic's top models may be temporary. The emergence of a cross-company standard for rating jailbreaks is a step toward large AI labs responding to security threats in a coordinated way rather than individually.

Frequently asked questions

Why was Claude Fable 5 unavailable?

Access to the model was limited by US export requirements; Anthropic restored it on July 1, 2026, after those restrictions were lifted.

What happens to requests that are considered suspicious?

The new cybersecurity classifier blocks the jailbreak technique from Amazon's report in more than 99% of cases and reroutes such requests to Opus 4.8 instead of Fable 5.