The U.S. government lifted the export controls that knocked Anthropic's Fable 5 and Mythos 5 offline for almost three weeks, and Fable 5 returned to users worldwide on July 1.
The fix was a more aggressive safety filter that blocks suspicious prompts, and government explicitly reserved the right to take the models away again.
The Trump administration lifted the export controls it had placed on Anthropic's two most powerful AI models, Fable 5 and Mythos 5, ending a nearly three-week standoff that knocked both models offline entirely. Fable 5 returned to users globally on July 1. Export controls are the rules the government normally uses to keep sensitive technology out of foreign hands, and applied here to an AI model.
The standoff began June 12, when the U.S. Commerce Department ordered Anthropic to block foreign nationals from using Fable 5 and the more powerful Mythos 5 over national security concerns. The trigger was a report from Amazon researchers who found a way to bypass Fable 5's safeguards, a technique known as a jailbreak, and get the model to identify software vulnerabilities. In one case, it produced code showing how to exploit them. Because the order took effect immediately and Anthropic had no reliable way to verify a user's nationality in real time, the company pulled both models offline for everyone.
The controls were lifted June 30 in a letter from Commerce Secretary Howard Lutnick, who wrote on X that the government had "worked closely with Anthropic to analyze and approve Fable 5." In exchange, Anthropic agreed to proactively detect and address security risks, help the government develop standards for future models, and flag any malicious activity it detects.
Fable 5 is now included in Pro, Max, Team, and select Enterprise plans for up to 50% of weekly usage limits through July 7. After that it moves to usage credits, a shift from the flat monthly subscription pricing most users plan around to pay-as-you-go metered billing.
SmarterX founder and CEO Paul Roetzer broke down what the resolution actually tells us, and what it doesn't, on Episode 224 of The Artificial Intelligence Show.
~3 weeks - Length of time the government took Fable 5 and Mythos 5 offline
2 - Commerce warning letters sent to Anthropic before the reversal
50% - Weekly usage limits that include Fable 5 through July 7
67% - Podcast listeners who want a light touch by government or none at all with model releases
18% - Favor the staggered, gated access that the government is now using
Fable 5 and Mythos 5 share the same underlying model. Fable 5, Roetzer explained, "is the Mythos model with safeguards." According to Anthropic's redeployment post, Mythos 5 can find and exploit software vulnerabilities more effectively than any other model, capabilities that make it "uniquely attractive to malicious actors." Fable 5 is the public version, launched with what Anthropic calls "the strongest safeguards we've ever applied to a model." The government decided those safeguards were not sufficient, and that's why the model was taken away.
The resolution arrived as a coordinated PR rollout. At 7:26 p.m. Eastern Time on June 30, White House chief of staff Susie Wiles, who does not post often on X, tweeted that "the United States is the undisputed winner in the AI race" and that "our shared priority remains get the best tech deployed as quickly and safely as possible." Lutnick retweeted her five minutes later. Anthropic's announcement came 20 minutes after that, followed by public thank-yous from co-founder Tom Brown, who had taken over the White House negotiations from CEO Dario Amodei, and researcher Sholto Douglas. "I read this as damage control," says Roetzer, noting that allies of the administration had been openly questioning its approach.
The fix comes down to a far more sensitive safety filter. The safeguards work through classifiers, software that reads each prompt and tries to judge whether the user's intent is harmful. If it decides yes, it can shut down the chat and even revoke access to the model. Anthropic's fix was to make those classifiers far easier to trigger, especially around areas such as biology and cybersecurity. That means more false positives: users with no harmful intent are already reporting blocked requests, which get rerouted to the less capable Claude Opus 4.8. "We just know they made the safeguards more likely to be triggered due to these classifiers being more aggressive," says Roetzer. "That is the gist of what we learned."
The deal covers one company, and even insiders call it opaque. Lutnick's letter binds only Anthropic. It says nothing about OpenAI, Google, or Meta, and it reserves the right to reimpose "a license requirement, should circumstances change or should Anthropic fail to adhere to its commitments." Dean Ball, an AI policy analyst who is joining OpenAI this week, wrote that "we have no idea what Anthropic did to make the models 'safe,' what commitments Anthropic has made going forward, and whether or how any of this applies to other frontier models in the government's licensing queue." His warning: "The opacity will not lend itself well to a stable, investable, trustworthy industry over time."
Roetzer's bigger concern is what this precedent means for every frontier model, the industry's term for the most advanced AI models available.
"If we get GPT 5.6 this week or if we get Gemini 3.5 Pro or whatever the next generations of frontier models are, assuming they're on par with the capabilities of Fable 5, Mythos 5, we're in essence entering a period where we don't even know what the government considers to be worthy of hitting these harmful intent triggers that shuts the model off or reroutes you to a weaker model."
— Paul Roetzer, founder and CEO of SmarterX, Episode 224 of The Artificial Intelligence Show
Two facts from this saga should affect how business leaders plan.
First, the government and Anthropic's trusted partners now have standing access to a more powerful model, Mythos 5, than the public can touch, and Roetzer expects that gap to be an expectation going forward.
Second, pre-release government review will likely slow how fast frontier models reach the rest of us, even if the labs' internal progress keeps accelerating. "They may be on GPT 6.5 by the time we get GPT 5.7," says Roetzer.
There is also a gap between what AI users want and what they just got. In the show's AI Pulse survey, 67% of listeners said the government should take a light touch or stay out of model releases entirely, while only 18% favored the staggered, gated access that is now the reality. For companies building on frontier models, the practical lesson is dependence: If your workflows run on a model Commerce can quickly switch off, availability becomes a risk you have to actively plan for.
GPT 5.6 is already in the government's licensing queue. According to Ball, it is fair to assume other model developers are in the early stages of submitting their models, too. The next frontier release will show whether the Anthropic agreement becomes the template for the whole industry or stays between one company and one administration.
Fable 5 could vanish as quickly as it returned. Roetzer noted that if someone finds a universal jailbreak or the new safeguards fail, the government has already signaled it can pull it back. That standing risk is exactly why he raised the question of whether more enterprises will move to open-source models they can train and control on their own, with their own data and their own application layer.
Frontier models are now core business infrastructure: 59% of organizations provide their employees ChatGPT, 46% Microsoft Copilot, 42% Google Gemini, and 37% Anthropic Claude, according to the 2026 State of AI for Business Report. The nearly three-week Fable 5 shutdown is a reminder that access to that infrastructure can change overnight, and most companies have not planned for that.
The full report, based on 2,100+ responses from professionals across roles, functions, and industries, maps which platforms organizations rely on, how far they have scaled, and whether they have the governance to adapt when the ground shifts. Read the full report →