Anthropic’s most powerful models are back. The U.S. Department of Commerce has lifted the export restrictions that cut off Claude Fable 5 and Mythos 5 from users around the world for two weeks. The company confirmed this shortly after Commerce Secretary Howard Lutnick announced on X that the models had been cleared for release.
The return of the Fable and Mythos models
Lutnick wrote that Anthropic would no longer need an export license. In return, the company pledged to proactively identify and address security vulnerabilities in its models, work with the government on standards for upcoming models, and report any potentially abusive activity. It also thanked users for their patience and everyone who helped bring the models back.
Mythos 5 remains restricted as it was initially. Anthropic received approval to restore it only for certain U.S. organizations, specifically entities that operate and protect critical infrastructure.
Fable 5 will return to the Claude platform, the Claude.ai app, and the Claude Code tool as early as Wednesday. Until July 7, it will also count as only half the usual usage toward the weekly limit for Pro, Max, Team, and some enterprise subscriptions.
Following talks with the government, Anthropic is deploying it with a new set of filters. These are intended to block cybersecurity-related tasks more consistently. This means that some routine activities, such as writing and debugging code, will now temporarily be handled by the less powerful Opus 4.8 model. The company promised to fine-tune the filters in the coming weeks so that they are less likely to mistakenly block legitimate requests and better distinguish actual abuse from legitimate work.
Anthropic is not tackling the issue alone. It is working with Amazon, Microsoft, Google, and other partners in Project Glasswing on a common approach to assessing attempts to bypass security safeguards. The joint framework is intended to determine how developers should respond to such findings.
Why the models were initially withdrawn
Let us return to the beginning of June, when Anthropic introduced Fable 5 as the first publicly available model in its top-tier Mythos family. Fable is essentially a version of Mythos with added safeguards so that the company can offer it to the general public. Anthropic described Mythos itself as the model with the world’s most powerful cybersecurity capabilities.
The excitement did not last long. Three days after launch, the company received a letter from the Department of Commerce. Citing national security powers, the government ordered Anthropic to cut off all foreign nationals from Fable 5 and Mythos 5, regardless of where they were located, including its own employees who were not U.S. citizens. Such an order cannot be enforced in real time across hundreds of millions of users. Anthropic therefore shut down both models entirely.
What triggered it? Someone found a way to bypass Fable’s security safeguards, a so-called jailbreak. According to reports, Amazon alerted the authorities after its researchers used a series of prompts to extract information from the model that could be used for cyberattacks and was supposed to remain off-limits. Anthropic, however, disputed the seriousness of the finding. It described it as a routine rather than universal method of bypassing security safeguards and argued that the same capabilities could be obtained from other publicly available models, including OpenAI’s GPT-5.5. The company then deployed a new safeguard intended to block the specific vulnerabilities reported by Amazon.
Tense relationship with the White House
Anthropic has been at odds with the Trump administration for much of the year. In February, federal agencies were instructed to stop using the company’s models. In March, the Pentagon designated Anthropic a supply chain risk, after which the company sued the Department of Defense. It had refused to work for the military without assurances that its tools would not be used for mass surveillance or autonomous weapons.
The ban also drew criticism from within the industry. Opponents warned that the U.S. government had handed extra time to Chinese open-source developers trying to close the West’s lead. Chinese company Zhipu is reportedly drawing closer to the best U.S. models just as Anthropic and OpenAI are being held back.
And it does not affect only Anthropic. OpenAI announced that it would initially release its new GPT-5.6 series only to a small group of vetted partners following government pressure for a gradual rollout.
What the experts say
The ban came at a sensitive time. Amazon is Anthropic’s largest investor, having invested roughly $13 billion in the company, while Anthropic is preparing to go public this year.
Experts are largely shrugging at the scale of the uproar over the bypassing of the safeguards. The lifting of the ban had been expected in technology circles, and reports about the Fable 5 jailbreak inflated the significance of the finding far beyond its actual importance. If such a reason were enough to block Fable and Mythos, competing models would reportedly have to be taken down as well.
The improvement in relations between Anthropic and the government is mainly attributed to company co-founder Tom Brown. The question remains whether the U.S. government will have to approve the release of every cutting-edge model. We are beginning to see this with the new GPT-5.6 series.
Sources: cnbc.com, fortune.com and aljazeera.com
