Claude Fable 5 and the Mythos class

Anthropic had previewed a model it called Mythos to a closed group of security organisations under a programme named Project Glasswing, because the model could find and fix software vulnerabilities at a level the company judged unsafe to release. On 9 June 2026 it released the same model to the public as Claude Fable 5, identical in capability but with classifiers that route requests touching cybersecurity, biology and the training of other models to a less capable model instead. Anthropic called it the most capable system it had ever made available. Mythos 5, without the safeguards, stayed with the Glasswing customers.

On 12 June the US Department of Commerce wrote to the company prohibiting access to either model for any non-American national on national-security grounds. Unable to enforce that at the level of individual users, Anthropic revoked access for every customer. Restoration began for American organisations on 26 June and the restrictions were lifted altogether on 30 June. It was the first time a government had pulled a deployed frontier model, and the first test of what the Action Plan's language about American leadership meant when the leading model was also the most dangerous one.

Fable 5.1 and Mythos 5.1 followed on 1 September 2026. The June episode is on this timeline as the point at which the capability the safety field had warned about, a model that could do serious harm in the wrong hands, stopped being hypothetical and became an export-control matter.