End of two-week blockade: Anthropic receives approval to export Mythos 5 to select partners in the US
The U.S. administration has officially lifted the ban on the export of Anthropic's most powerful AI model, Claude Mythos 5. This decision, signed by Secretary of Commerce Howard Lutnick, opens access to the model for more than 100 American organizations, including leading corporations and government agencies. This refers to companies and agencies from the so-called Appendix A — a list of trusted partners for which a separate license to obtain the model is no longer required.
This step effectively ends a two-week standoff between the White House and Anthropic, which arose from concerns that advanced technology could fall into the wrong hands. Previously, as we recall, both of the company's models — Mythos and Fable — were disabled after Anthropic's key investor, Amazon, expressed serious concerns about their safety. The company's researchers warned that Fable 5, in particular, could be hacked for potentially dangerous purposes.
Mythos 5: From Secret Testing to Widespread Deployment
Until now, Mythos was used exclusively within the Glasswing program — a vulnerability research project involving about 150 organizations from 15 countries. The results were impressive: within hours, the model identified critical errors in classified systems during government testing. Now, after the lifting of restrictions, Mythos 5 will be able to apply its analytical capabilities to strengthen the cybersecurity of key U.S. structures.
Fable 5 Still Pending: A New Era of AI Regulation
The fate of the public version of Fable 5 remains uncertain. Unlike Mythos, access to which was initially restricted, Fable 5 was previously open to all subscribers and was considered the most powerful publicly available AI tool. Its launch is now postponed indefinitely. The situation is compounded by a new executive order signed on June 2, which introduces a voluntary federal review channel for advanced models. Developers can now submit their solutions for a 30-day cybersecurity check before launch.
Notably, OpenAI has followed the same path, restricting access to the most powerful version of GPT-5.6, called Sol, to approximately 20 state-approved partners. The weaker versions, Terra and Luna, have become available to the general public.
Cryptalist Analysis: We are witnessing the formation of a new two-tier system of access to AI. On one hand, there are "elite" models for selected partners; on the other, public but limited versions. This creates a precedent where not only chips but also the algorithms themselves become tools of geopolitical control. The question of whether Fable 5 will receive similar approval will serve as a litmus test for Washington's entire new AI policy.