Anthropic Allowed to Release Fable 5, Must Cooperate with U.S. Government
Following a brief suspension due to US export controls, Anthropic has announced the phased redeployment of its latest AI models. While the “Fable 5” model will be made globally available, access to “Mythos 5” will remain strictly regulated for the time being.
After access to the new models Claude Fable 5 and Claude Mythos 5 was interrupted on June 12 due to US export controls and difficulties in verifying user nationality, Anthropic has presented a plan for resuming operations. The return of the models is tied to specific usage requirements, new security mechanisms, and closer cooperation with government agencies.
Availability and Cost Models
The distribution of the two models follows different strategies:
Claude Fable 5 (Global Rollout):
Starting July 1, 2026, Fable 5 will be available worldwide via the Claude platform, Claude.ai, Claude Code, and Claude Cowork. The following terms apply for billing and usage:
- Pro, Max, Team, and select Enterprise plans: These users will receive a quota covering up to 50% of their weekly usage limits until July 7.
- From July 8: For these subscribers, the use of Fable 5 will be handled via a system of usage credits.
- Cloud Providers: Integration into AWS, Google Cloud, and Microsoft Foundry will be implemented “as quickly as possible,” according to Anthropic.
Claude Mythos 5 (Restricted Access):
In contrast to Fable 5, Mythos 5—a model with specialized cybersecurity capabilities—will initially only be re-enabled for a specific set of US organizations. Further access will be controlled through the “Glasswing program” in coordination with the US government.
Conditions: New Safety Standards and “False Positives”
The return of the models is largely contingent on the implementation of improved safety measures. The background of the suspension was a report from Amazon researchers indicating that Fable 5’s safety barriers could be bypassed through targeted prompting to identify software vulnerabilities.
As a condition for the restart, Anthropic has implemented the following measures:
- Improved Safety Classifiers: A new AI-based filter has been trained that blocks the described bypass technique in over 99% of cases.
- Increased Safety Margin: To minimize risk, the filters have been set to intervene even at ambiguous requests.
- Functional Costs of Safety: Anthropic admits that these stricter filters represent a “technical cost” for users: there will be a higher rate of so-called “false positives.” This means that legitimate requests during routine coding and debugging tasks will more frequently be blocked in error.
Industry Framework and Government Cooperation
Another central aspect of the return is the commitment to closer cooperation with both the government and competitors. Anthropic is working together with partners such as Amazon, Microsoft, and Google to develop a unified industry framework for objectively assessing the severity of “jailbreaks” (attempts to bypass safety barriers).
Furthermore, the company is committing to deeper collaboration with US authorities. This includes providing government partners with early access to new models for testing, accelerated information sharing regarding newly discovered security vulnerabilities, and dedicated research resources in the field of AI security.
