Anthropic has secured permission to restore Mythos 5 access for a limited set of trusted customers after U.S. officials forced the company to cut off foreign users from its most advanced models. The partial reversal is narrow, but it signals that Washington is willing to trade blanket restrictions for a more selective model of control when the security case is strong enough.
The company had been unable to block access only for specific users, so it disabled almost everyone from using Mythos 5 and Fable 5 for cyber defense testing. Now, after talks with the U.S. government, Anthropic says a small group of cybersecurity specialists and infrastructure providers can use Mythos 5 again, while it continues pushing to bring Fable 5 back into general use.
U.S. government approves limited Mythos 5 access
Bloomberg reported that the move was cleared with the U.S. government. In a letter to Anthropic’s head of compute, Commerce Secretary Howard Lutnick said the company had worked with officials to reduce the risks linked to the models and that ”substantial progress” had been made. He added that access could be restored for certain trusted partners.
A separate Commerce Department official, Benno Kaas, said the work was done in two weeks and was intended to keep the U.S. ahead in AI without weakening national security. That is the polite version of a familiar Washington bargain: move fast, but not so fast that the lawyers panic.
Mythos 5 comes back first
Anthropic has not said that Fable 5 gets the same treatment yet. The government letter does not mention any change for that model, even though Anthropic says it will keep working over the weekend to restore broader access and set up a framework for similar cases later.
- Mythos 5 access: restored for a small group of cyber defense specialists and infrastructure providers
- Fable 5 access: still not addressed by the U.S. letter
- Formal access base: about 200 companies and organizations had qualified for Mythos 5 at the start of the month
OpenAI had to play by similar rules
The episode puts Anthropic in the same regulatory bucket as OpenAI, which had to clear customers with U.S. authorities before distributing GPT 5.6. That is a reminder that frontier AI companies are no longer just shipping software; they are also negotiating the terms of access with governments that see advanced models as dual-use tools.
Sources familiar with the talks say CEO Dario Amodei stayed out of the spotlight to avoid irritating officials, with whom he reportedly has a complicated relationship. That may have been a smart bit of restraint: in Washington, sometimes the loudest person loses the room. The bigger question now is whether this selective reopening becomes the template for every future model release, or whether the next security scare shuts the door again.