Anthropic's Ghost Model Trains in the Shadows While Export Drama Drags On 🕵️

Anthropic has completed training on a more capable successor to its suspended Mythos 5 model, according to AI commentator Andrew Curran, though the company has not confirmed the system's existence, name, or capabilities. Curran said the model could ship as Mythos 5.1 or Mythos 6, or remain internal, in a post made nine days after US export controls forced Anthropic to disable Mythos 5 and Fable 5. "A new, more capable version of Mythos has emerged from training. I don't know whether it will be called Mythos 5.1 or Mythos 6, or if Anthropic will keep it internal to accelerate further development," Curran wrote.

Anthropic released Mythos 5 and Fable 5 on June 9, and three days later Commerce Secretary Howard Lutnick sent a directive to Anthropic chief Dario Amodei citing national security. The order barred every foreign national, including Anthropic's own foreign-born staff, prompting the company to disable both models worldwide. The government flagged a method for bypassing Fable 5's safeguards; Anthropic reviewed the demonstration, called it narrow, and warned that applying the same standard would halt new model launches industrywide. Fable 5 ships with heavy safeguards for public use, while Mythos 5 runs with fewer security restrictions through Project Glasswing, the firm's vetted cybersecurity program. Anthropic says about 50 partners have used early Mythos models to find more than 10,000 high- or critical-severity software flaws.

The suspension followed a reported Amazon warning to officials. Amazon has committed up to $25 billion to Anthropic, but reportedly told the administration that its researchers used Fable 5 to surface attack-ready information. A bipartisan group of lawmakers is pressing Lutnick to explain the legal and technical rationales behind the export controls, and the clash drew public comments from Donald Trump and a defense from Amodei. Anthropic is still working to reverse the restrictions.

Curran argued that halting deployment does little to slow progress and may even accelerate it by freeing resources, pointing to open-weights rivals such as Z.ai's GLM-5.2, which he said matches far larger closed models on coding tasks at a fraction of the cost. Whether the trained successor ships publicly, stays restricted to Glasswing partners, or remains internal for further development remains unconfirmed.