Claude Fable 5 returns worldwide as US lifts export controls

The availability of frontier artificial intelligence models is no longer solely determined by technological readiness or corporate strategy. It is now subject to the complex dynamics of international policy and government regulation. On 1 July 2026, Anthropic restored global access to Claude Fable 5 and Mythos 5, following the lifting of a US Department of Commerce export control order that had forced the company to suspend the models worldwide just days after their initial launch.

This marks a significant moment in the governance of artificial intelligence. The temporary suspension of these models underscores how quickly regulatory bodies can intervene when national security concerns are raised. For project delivery professionals and enterprises relying on these advanced tools, the episode serves as a stark reminder that standardising operations on a single vendor carries new types of geopolitical risk.

Anthropic initially launched Fable 5 and Mythos 5 on 9 June 2026. Fable 5 was designed for general use with robust safety mechanisms, while Mythos 5, possessing fewer safeguards, was restricted to a select group of trusted partners for defensive cybersecurity applications.

Just three days later, on 12 June, the US government issued an export control directive citing national security authorities. The directive required Anthropic to suspend access for all foreign nationals. Given the technical challenges of verifying user nationality in real time, the company opted to disable the models globally to ensure compliance.

The government's action was prompted by a report from Amazon researchers who had discovered a method to bypass Fable 5's safeguards. The researchers prompted the model to identify software vulnerabilities and, in one instance, generated code demonstrating how a vulnerability could be exploited.

During the suspension period, Anthropic worked closely with the US government and its partners to address the issues raised in the report. The company's testing revealed that the reported technique did not expose any unique, highly dangerous capabilities exclusive to Fable 5. In fact, several other less capable models from various developers could identify the same vulnerabilities and produce similar exploit demonstrations.

The behaviour in question fell into a grey area for Fable 5's safeguards. The model had been programmed with a wide "safety margin," meaning it was designed to block requests that were likely benign but carried a small risk of being harmful. The reported bypass allowed access to routine defensive cybersecurity tasks rather than novel offensive capabilities.

Nevertheless, Anthropic implemented an improved safety classifier specifically targeted at blocking the behaviour detailed in the Amazon report. If a user now submits a request that triggers this classifier, the request is blocked and routed to the older Opus 4.8 model instead.

With the new safeguards in place and the export controls lifted, Anthropic redeployed Fable 5 across its platforms. The company introduced temporary usage caps for certain subscription tiers:

• Pro, Max, Team and Enterprise plans: Limited to 50% of weekly usage allowance until 7 July

• After 7 July: Access managed via usage credits at standard rates

• Cloud platforms: Re-enabled access via AWS, Google Cloud and Microsoft Foundry as quickly as possible

• Mythos 5 access: Restored for approved US organisations, with ongoing coordination to expand access through the Glasswing programme

By rolling out the model in phases, Anthropic can manage demand while validating its new safeguards across different deployment scenarios.

The events surrounding the suspension and redeployment of Fable 5 have highlighted a critical gap in the artificial intelligence industry: the lack of a standardised method for assessing and categorising "jailbreaks" or techniques used to bypass model safeguards.

Without a shared framework, it is difficult for developers, researchers and government officials to communicate the severity of a vulnerability consistently. A minor bypass that allows a model to perform a routine, non-harmful task might be treated with the same urgency as a critical flaw that exposes dangerous capabilities.

To address this, Anthropic has initiated efforts to develop a consensus framework in collaboration with other major technology companies, including Amazon, Microsoft and Google. Establishing a common standard will help the industry triage new findings more effectively and provide clearer guidance to regulatory bodies.

• The temporary suspension of Fable 5 demonstrates that government export controls can disrupt global access to frontier models almost instantly.

• Organisations standardising on specific artificial intelligence platforms must account for geopolitical and regulatory risks in their procurement strategies.

• The implementation of stricter safety classifiers may result in a higher rate of benign requests being flagged or routed to older models.

• The industry lacks a standardised framework for assessing the severity of model jailbreaks, complicating communication between developers and regulators.

• Close collaboration between technology companies and government agencies is becoming a prerequisite for the deployment of advanced capabilities.

To explore the implications of AI governance and export controls on project delivery, subscribe to the Project Flux newsletter for ongoing analysis.

All content reflects our personal views and is not intended as professional advice or to represent any organisation.