Claude Fable 5 is back after 19 days as US export controls lift

Just three days after Anthropic publicly released what it called its most powerful AI yet, Claude Fable 5 and its higher-tier sibling Mythos 5, both models went dark worldwide on June 12. The reason was neither a performance problem nor an outage. The U.S. government issued a directive ordering Anthropic to suspend access, and the company complied by disabling the two models itself (Anthropic statement).

The directive arrived at 5:21 p.m. ET on June 12 (6:21 a.m. JST on June 13). By that evening, Anthropic had cut off Fable 5 and Mythos 5 for every customer worldwide, paying users included. Other Claude models such as Opus and Sonnet kept working normally, so only the two models billed as the strongest were pulled.

Three days earlier, we covered the launch of these two models in "Why Engineers Cheered, Then Revolted." The sequel turned out to be "pulled by the government within 72 hours, back within 19 days." What follows lays out what happened, why the whole world was caught in the net, and what came attached to the lifting.

A 19-Day Shutdown, and How It Came Undone

The controls did not lift all at once. On June 26, Commerce Secretary Howard Lutnick sent a second letter. It stated that a license would no longer be required to export, reexport, or transfer Mythos 5 in-country to the entities listed in Annex A — the "trusted partners" — to those entities' foreign-national employees, or to Anthropic's own foreign-national employees. The same day, restored Mythos 5 access for a set of U.S. organizations was approved. Everyone not on the list still needed a license, making this a narrow carve-out rather than a reversal (Tech Policy Press).

Four days later, on June 30, the Commerce Department lifted the export controls on Fable 5 and Mythos 5 entirely. Anthropic brought Fable 5 back worldwide the next day, July 1, across the Claude Platform, Claude.ai, Claude Code, and Claude Cowork, and said it would re-enable access on AWS, Google Cloud, and Microsoft Foundry as quickly as possible.

Usage terms got a transition period. On Pro, Max, Team, and select Enterprise plans, Fable 5 was included for up to 50% of weekly usage limits through July 7, after which availability shifted to usage credits. Japanese outlets Impress Watch and ITmedia AI+ both reported the restart on July 1.

One detail is easy to miss: Mythos 5 is not back to where it was. It was restored for an approved set of U.S. organizations only, and Anthropic says it continues to coordinate with the government to widen access to the broader domestic and international partners in Project Glasswing, the limited program it started in April. As of July, the model that returned to the whole world is Fable 5 alone.

Three Days After Launch, the Most Powerful AI Went Dark Everywhere

The trigger was an export control directive from the U.S. Commerce Department, the agency that oversees trade and exports. Issued in the name of Commerce Secretary Howard Lutnick, it ordered Anthropic to suspend access to Fable 5 and Mythos 5 for any foreign national (9to5Mac). It applies whether the person is inside or outside the United States, and it even includes Anthropic's own foreign-national employees.

The problem is that an online service cannot reliably sort users into "foreign national" and "U.S. citizen" in real time on every request. To be sure it complied, Anthropic chose not to block only foreign nationals but to disable both models for all customers at once. The company said it "must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance."

In Japan, Impress Watch and ITmedia reported it the same day, while abroad NBC News and Bloomberg Law picked it up at once. An AI company yanking its flagship product worldwide on a single government order is a rare event.

The Trigger Was a Claimed "Jailbreak" — by Amazon

The immediate spark was a company claiming it could jailbreak Fable 5. A jailbreak is a trick that slips past an AI's safety guardrails to pull out dangerous content the model is supposed to refuse. At the time of the shutdown the claimant was described only as "another company"; later reporting identified it as researchers at Amazon, who found a way around the guardrails by getting Fable 5 to identify software vulnerabilities — which is what prompted White House concern (CSIS). Because Fable 5 and Mythos 5 are built on the model Anthropic itself sealed off in April as "too dangerous to release" over its hacking ability, a broken guardrail was always going to make the government nervous.

Anthropic, however, sharply disputes what the jailbreak actually showed. When it reviewed the demonstration, it found only "a small number of previously known, minor vulnerabilities" — and ones that "other publicly-available models are able to discover without requiring a bypass." It said the same behavior could be reproduced on many far less capable models, naming OpenAI's GPT-5.5 as one example (StartupHub.ai).

Anthropic also maintains that Fable 5's safeguards are "substantially more effective than those of any previously deployed model." In short, its position is that the supposed hole is small and common to rival models, so using it to recall a commercial product used by hundreds of millions makes no sense. The redeployed version shipped with tightened safeguards anyway: a classifier that catches requests headed toward vulnerability exploitation and routes them to the less capable Opus 4.8. Anthropic is now working with Amazon, Microsoft, and Google on an industry standard for grading how severe a given jailbreak actually is.

Anthropic's Rebuttal: "This Standard Would Halt Every New Model"

Anthropic is complying while clearly objecting. The sharpest line in its statement is its warning about the knock-on effect across the industry: "If this standard was applied across the industry, we believe it would essentially halt all new model deployments."

Any large AI, if you look hard enough, has small holes. If "one minor jailbreak found means a worldwide shutdown" were applied for real, OpenAI, Google, and Anthropic would all hit the same wall with every new release. So Anthropic treated the call as "likely based on a misunderstanding" and closed by pledging to "work toward restoring access as soon as possible." The talks with the government took about two and a half weeks to land.

This is not the first time the safety-first company has clashed with the government over safety. Anthropic has previously fallen out with Washington on safety grounds, and it has staked out a "safety-first" identity in other ways too, such as declaring it would never put ads in Claude. This time, that very banner ended up tying its own hands.

What Happened in Four Months, on a Timeline

To make sense of the episode, it helps to trace the past four months. From the moment Mythos surfaced, through the sealing, the launch, the cutoff, and the restart, events moved remarkably fast.

← swipe to move

The Model It Sealed as "Too Dangerous" Was Stopped by the Government

There is a strange symmetry here. Mythos, the base for Fable 5 and Mythos 5, was the model Anthropic itself sealed in April as "too dangerous to release" over its hacking ability (Just Security). It then released that model on June 9 on the grounds that "safeguards make it safe now" — and three days later, the government stopped it on the grounds that it is, after all, dangerous.

The April sealing was already a national-level tug-of-war: the White House opposed wider access on security grounds (Axios), and in Japan, Mythos came up at an LDP cybersecurity strategy meeting in April. Anthropic urged the industry to "hit the brakes," lowered its most powerful model to the public anyway, and then had the brakes hit on it by the government. A company that flies the safety flag now finds itself pinned down by a safety call.

Anthropic is no stranger to things spilling out in unexpected ways, from the time 510,000 lines of Claude Code source code leaked. Once again, it has had to pull its own freshly launched centerpiece off the shelf.

Why Some People Say "the Shutdown Was Staged"

You cannot discuss this shutdown without the event that came about ten days earlier. On June 4, Anthropic published a blog post urging the whole industry to voluntarily pause work on frontier AI (Fortune). CEO Dario Amodei named three concerns: AI-made bioweapons, mass surveillance by authoritarian states, and "recursive self-improvement," where AI builds its own next version without humans watching. In short, his worry was that defense and society's safeguards are not keeping up with how fast AI's capabilities (the offense) are growing — a concern many experts genuinely share.

The catch is who was saying it: a player widely seen as having pulled clearly ahead in the performance race. When Anthropic tells everyone else to slow down, a familiar criticism follows — that it is knee-capping rivals under the banner of safety. AI policy figure David Sacks and some White House officials have accused the company of regulatory capture: overstating risk to slow competitors. Strict compliance regimes, the argument goes, are easiest for the largest incumbents — who already run red teams (offensive test units) and documentation pipelines — to absorb, while acting as a barrier, a "moat," against startups and open source (The Dossier). OpenAI's Sam Altman once branded this kind of pitch "fear-based marketing."

Lay this shutdown over that context and one reading becomes unavoidable: that the most powerful model going dark three days after launch was itself on script — a foregone conclusion that fits the slow-down narrative. Each dramatic move — seal, release, suspend — deepens the sense that Anthropic's products are "so dangerous and so special" that "regulation is therefore needed," a current that, conveniently, runs in the company's favor.

That said, this is speculation, and there is a problem with it. The party that ordered the shutdown was the U.S. government, and Anthropic itself objected, calling the directive "likely a misunderstanding" and spending 19 days pushing until it was withdrawn. If it were a staged script, there would be little reason to fight it. "The pause call is a fair point" and "the pause call happens to suit Anthropic competitively" can both be true at once. The same event looks completely different depending on which lens you use — and that very difficulty captures the position of a company that flies the safety flag while also leading on raw capability. Amodei has consistently sounded the alarm, including in his own essays, but whether his words are taken at face value is now a separate question.

What It Means for Users and Companies in Japan

This was not a faraway story. The June 12 directive targeted "any foreign national." From the U.S. point of view, users and companies in Japan are foreign nationals. So Japanese engineers and businesses that had only just started using Fable 5 on June 9 were caught in the cutoff too. Even on a paid plan, they woke up one morning to find that only the most powerful model was no longer selectable — and it stayed that way for 19 days.

Another angle that matters is the impact on cyber defense. As we reported at launch, in the months after April, the Japanese government and major Japanese banks had moved to obtain access to Mythos for cyber defense. Countries were effectively fighting over "who gets to use" a model billed as having the strongest defensive ability. For those who were counting on that capability, a trump card they had just acquired — or were about to — went dark on a single U.S. directive. And what came back worldwide on July 1 was Fable 5; Mythos 5 was restored only for an approved set of U.S. organizations. Seen from Japan, the door to the top-tier model is not fully open.

Daiwa Institute of Research framed the episode, in a July 3 report, as a demonstration of the limits of AI regulation. The capabilities that alarmed regulators in Fable 5 are reproducible in rival models and in Chinese open-weight models, which makes supply-side controls on one model inefficient. Regulation should shift from "stopping" a specific model toward managing proliferation as a given, and Japan should join the work on safety-evaluation standards and diversify its AI procurement to preserve room for "sovereign AI" (Daiwa Institute of Research, Japanese).

Throughout the outage, models long used in production such as Opus and Sonnet kept running, so day-to-day work never ground to a halt. But the episode makes one fact plain: cutting-edge American AI can be cut off at any time by a single U.S. security decision, regardless of your contract or what you paid. Coming back after 19 days does not remove that mechanism. It is worth filing away as a geopolitical risk before you build your operations deeply on any one service.

Why It Became a Worldwide, All-at-Once Shutdown

The directive itself asks for "suspension for foreign nationals," so U.S. citizens are technically not the target. The reason the whole world still went dark lies in how the models are served. Claude is used through a browser chat (Claude.ai), an API that developers call from their programs, the coding tool Claude Code, and more. Building a system to accurately judge each user's nationality on every request and route them instantly is not realistic on short notice.

So, to avoid the risk of "blocking only some and missing others," Anthropic took the safe side and shut everyone off. In export-control terms, there is a concept that letting a foreign national merely touch software or technical data counts as an "export." That is why the order reaches even foreign-national employees inside the company. You can read this as cutting-edge AI starting to be treated like semiconductors or cryptography: a strategic good that must not cross borders.

If this line holds, the impact will not stop at these two models. As Anthropic warned, if "one minor jailbreak means a worldwide halt" becomes routine, every AI company carries the same risk with each new release. CSIS flagged a deeper ambiguity: Commerce imposed the controls by letter rather than through formal rulemaking, and it remains unsettled whether remote access to a hosted model even constitutes an "export" under existing law (CSIS). The controls lifted; that ambiguity did not.

What Was "Still Unsettled" — and How It Settled

When we published this on June 13, three questions were open. Within 19 days, all three had answers. Here is what we were waiting on, and where each one landed.

What the 19 Days Left Behind

The controls are gone, but the episode did not simply erase itself. Three things stayed.

First, the commitments attached to the lifting. Anthropic agreed that for models materially advancing the capability frontier in areas relevant to national security, it will give designated government partners expanded early access to both the models and their safeguards; that when significant jailbreaks or misuse patterns are identified it will investigate, triage, and notify government counterparts quickly; and that it will work with the government and industry peers toward a shared, voluntary security and evaluation standard for frontier model providers. Before the shutdown the relationship was "the government issues a directive, the company complies." After the restart, there is a standing channel: show it early, report what breaks, write the rules together. Anthropic itself accepted the government moving structurally inside its safety decisions.

Second, the European reaction. Per CSIS, European politicians cited the controls as evidence that dependence on U.S. models is a supply chain vulnerability, sharpening the push for "sovereign AI." Coming back in 19 days did less to weaken that argument than to give it a number: a foreign government's decision can switch off a production model for nineteen days. Daiwa Institute of Research reached a parallel conclusion for Japan, urging diversified procurement to preserve autonomy.

Third, Mythos 5 has not fully come back. What returned to the world is Fable 5; the higher-tier model remains limited to approved U.S. organizations while Anthropic works with the government to widen it to Glasswing's domestic and international partners. For the most capable model, a mechanism where the U.S. government decides case by case who gets access is still in place. We have followed the Mythos saga from our launch coverage onward, and we will update this article as Mythos 5 access widens.

Sources

• ▸Anthropic - Redeploying Claude Fable 5 (June 30, 2026)
• ▸Anthropic - Statement on the US government directive to suspend access to Fable 5 and Mythos 5 (June 12, 2026)
• ▸Tech Policy Press - Commerce Eased Its Block on Anthropic's Mythos, But Major Questions Remain (June 27, 2026)
• ▸CSIS - The Department of Commerce Restricted Access to Anthropic's Latest Models. What Comes Next? (June 16, 2026)
• ▸Impress Watch - Anthropic's Claude Fable 5 returns July 1 as export controls are lifted (Japanese) (July 1, 2026)
• ▸ITmedia AI+ - US government lifts export controls on Fable 5 / Mythos 5 (Japanese) (July 1, 2026)
• ▸Daiwa Institute of Research - The limits of AI regulation shown by Fable 5's return (Japanese) (July 3, 2026)
• ▸9to5Mac - Anthropic pulls Claude Mythos 5 and Claude Fable 5 following US government directive
• ▸NBC News - Anthropic suspends new AI models after government directive
• ▸Bloomberg Law - Anthropic Says US Limits Foreign Access to Fable 5, Mythos 5
• ▸StartupHub.ai - Anthropic Disables Fable 5 and Mythos 5 After US Export Control Order
• ▸Impress Watch - Anthropic suspends Mythos-class models on US government order (Japanese)
• ▸ITmedia NEWS - Fable 5 and Mythos 5 fully suspended on US government order (Japanese) (June 13, 2026)
• ▸Fortune - Anthropic warns AI could soon build itself, urges a global pause on development (June 5, 2026)
• ▸The Dossier - The Federal AI Moat: Anthropic Is Selling Regulatory Capture as "Bipartisan Wisdom"
• ▸kkm-mako - Claude Fable 5 Goes Public: Why Engineers Cheered, Then Revolted