Fable 5 Came Back Under a Commerce On-Off Switch

The U.S. Commerce Department lifted export controls on Anthropic’s Claude Fable 5 and Mythos 5 on June 30, 2026, 18 days after a national-security order took both models offline worldwide. Anthropic restored Fable 5 for global users the next day, on July 1, after training a new safety classifier with federal testers.

Access came back. The Commerce letter that removed it is now the practical gate for frontier releases, including OpenAI’s GPT-5.6 preview that stayed closed until July 9.

Eighteen Days Offline, Then a Letter From Lutnick

Anthropic launched Fable 5 and Mythos 5 on June 9, 2026, three days before the cutoff. Fable 5 was the public, safeguarded build. Mythos 5 was the same underlying system with some of those safeguards lifted, limited at launch to Project Glasswing cyber defenders working with the U.S. government.

THE 18-DAY FABLE 5 BLACKOUT

1. June 2, 2026: President Donald Trump signs Executive Order 14409, a voluntary 30-day pre-release review for covered frontier models.
2. June 9, 2026: Anthropic launches Claude Fable 5 for general use and Claude Mythos 5 for Glasswing partners.
3. June 12, 2026: At 5:21 p.m. Eastern, Commerce orders Anthropic to block foreign nationals, including its own non-U.S. staff. Both models go dark for every customer that night.
4. June 26, 2026: A follow-up letter clears Mythos 5 for a set of U.S. organizations. OpenAI starts a limited GPT-5.6 preview for a small group of trusted partners.
5. June 30, 2026: Commerce Secretary Howard Lutnick withdraws the license requirement for both Anthropic models.
6. July 1, 2026: Fable 5 returns globally on Claude.ai, Claude Code, Claude Cowork, and the Claude Platform.
7. July 9, 2026: GPT-5.6 Sol, Terra, and Luna go public after the preview.
8. September 1, 2026: Anthropic ships Fable 5.1 for general use and Mythos 5.1 inside trusted-access programs.

Lutnick said his department had worked with Anthropic over those two weeks to analyze and approve Fable 5. His June 30 letter also said the administration could put the controls back if circumstances changed or the company failed its commitments.

The Foreign-Person Rule Pulled the Plug for Everyone

The June 12 directive did not say “turn the models off for Americans.” It told Anthropic to suspend all access by foreign nationals, whether they sat inside the United States or outside it, and it named Anthropic employees who are not U.S. citizens. Anthropic said it had no reliable way to check nationality in real time, so it disabled Fable 5 and Mythos 5 for all users to stay inside the order. Other Claude models kept running.

The US government, citing national security authorities, has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees.

The net effect of…

— Anthropic (@AnthropicAI) June 13, 2026

That is how an export rule aimed at foreign persons became a worldwide outage. Under a deemed export, giving a controlled technology to a foreign national on U.S. soil counts as an export to that person’s home country. A German engineer in a California office would have been an export. An API that cannot read passports cannot draw that line, so the compliant move was a total cutoff. Live Fable 5 sessions died with an error the same evening. New calls fell back to a user’s default model or to Claude Opus 4.8.

WHAT A DEEMED EXPORT MEANS HERE

• The target: Foreign persons, including staff and users inside the United States, not a blanket order to shut out U.S. citizens.
• The legal reach: A deemed export treats in-country access by a foreign national as an export to that person’s home country.
• The operational bind: Anthropic could not verify nationality across its products in real time, so it took both models down for everyone.

Anthropic called the underlying finding a narrow jailbreak and said it had reviewed a demo that surfaced a small number of previously known, minor software bugs. The company said it would comply and that it disagreed with using that finding to recall a commercial model.

We disagree that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people.

Anthropic, June 12, 2026 statement

Amazon researchers had shown a prompt path that got Fable 5 to identify vulnerabilities and, in one case, produce exploit code. Anthropic later said the technique did not unlock unique Mythos-level offensive skills and that it sat in a borderline band its classifiers already treated with extra caution.

Washington’s Own Order Barred a Mandatory License

Ten days before the cutoff, the White House had told agencies to build a voluntary path, not a permit. Executive Order 14409, titled Promoting Advanced Artificial Intelligence Innovation and Security, asked developers to share covered frontier models with the federal government for up to 30 days before other trusted partners saw them. Section 3 also said the order should not be read as creating no mandatory governmental licensing, preclearance, or permitting for new models.

Nothing in this section shall be construed to authorize the creation of a mandatory governmental licensing, preclearance, or permitting requirement for the development, publication, release, or distribution of new AI models, including frontier models.

Executive Order 14409, White House, June 2, 2026

Commerce still used an export-control letter, with criminal and civil penalties attached, as the enforcement tool. The June 26 Mythos letter named the organizations that could use the model and reserved the right to change that list at any time. The June 30 lift withdrew the license requirement after Anthropic agreed to detect security risks, work on future-release protocols, and report malicious activity. The voluntary order and the letter now sit side by side. Labs that want a wide launch negotiate the letter.

GPT-5.6 Sat in Preview Until July 9

OpenAI hit the same gate from the other direction. The company previewed GPT-5.6 on June 26 as three models, Sol, Terra, and Luna, and held the public rollout at the administration’s request. OpenAI CEO Sam Altman said, “This isn’t quite the process that we think is optimal.” The company wrote that it did not believe this kind of government access process should become the long-term default, and that it was taking the short-term step to get broader availability.

The preview lasted until July 9, 13 days after it opened, when Sol, Terra, and Luna went public. That was a cleaner path than Anthropic’s, because OpenAI never had to kill a live consumer model overnight. It was the same gate. Washington saw the model first, named the first customers, and then allowed a wider release. Altman’s complaint was about that sequence, not about a single Anthropic feud.

What Came Back on July 1, and What Did Not

On July 1 Anthropic started the Fable 5 and Mythos 5 return across Claude.ai, Claude Code, Claude Cowork, and the Claude Platform. The company said the export controls have been lifted as of June 30. Fable 5 came back for users globally. For Pro, Max, Team, and select Enterprise plans it counted toward up to 50% of weekly usage limits through July 7, then moved to usage credits. Anthropic said it would re-enable the model on AWS, Google Cloud, and Microsoft Foundry as quickly as possible.

WHO GOT WHICH MODEL AFTER THE LIFT

The technical fix was a retrained safety classifier aimed at the Amazon technique. Anthropic said that classifier now blocks the reported bypass in over 99% of cases. When it fires, the user is told and the request is sent to Opus 4.8. Researchers at the Commerce Department’s Center for AI Standards and Innovation tested the old and new safeguards and, Anthropic said, agreed they were extraordinarily strong. The tradeoff was more blocked coding and debugging requests. That is the friction that survived the lift: Fable 5 was back, and more ordinary work bounced to a weaker model.

Glasswing Holds the Unrestricted Model

Mythos 5 did not become a public product on July 1. Anthropic restored it for the U.S. organizations approved on June 26 and said it was still working with the government to widen Glasswing to more domestic and international partners. The unrestricted cyber build stayed on an allowlist even after the license requirement fell away. That split is the product design, and it is also the political settlement.

WHAT ANTHROPIC AGREED TO FOR THE LIFT

• New classifier: A safeguard trained with the government to block the Amazon jailbreak path and route flagged Fable 5 traffic to Opus 4.8.
• Malicious-activity reports: A pledge to tell the government when it finds abuse involving the models.
• Future-release protocols: Deeper pre-release testing, information sharing, and research work with U.S. agencies.
• Jailbreak scoring: A draft industry standard with Amazon, Microsoft, Google, and other Glasswing partners, scoring capability gain, breadth, ease of weaponization, and discoverability.

Those terms are how a lab gets the letter withdrawn. They are also how Washington stays inside the loop on the next build. Glasswing began in April as the trusted path for Mythos-class cyber tools. The June fight did not end that path. It made the public model the one with the heavy classifiers, and it left the stronger system with the partners the government had already accepted.

September’s 5.1 Release Runs on the Same Split

On September 1, 2026, Anthropic released Claude Fable 5.1 and Claude Mythos 5.1. They are again the same model with different safeguards. Fable 5.1 is generally available for Pro, Max, Team, and Enterprise users and on AWS, Google Cloud, and Microsoft Foundry. Mythos 5.1 stays inside trusted-access programs for vetted cyber defenders and life scientists. A biology access program, built with the U.S. government, is meant to open enrollment for scientists after an invite-only beta.

The new public build is cheaper on typical token bills by about 25%, and Anthropic said highly agentic jobs can save up to about 45% because cache-read prices fell. Cyber classifiers, the company said, now block 60% fewer false positives than before, in part because Fable 5.1 may identify software vulnerabilities in source code but still may not write exploits. Biology safeguards, it said, now intervene on benign requests 85% less often than the ones that shipped with Fable 5. Dual-use biology and chemistry questions still go to Opus models. Penetration testing, exploit generation, and binary-based vulnerability scanning stay blocked on the public build.

That is the June settlement, version 5.1. The general model is stronger and less twitchy. The unrestricted model is still a government-coordinated club. Lutnick’s reservation of rights was never withdrawn in public, and the next letter would travel the same route as the last one.

Frequently Asked Questions

Why Did the US Government Suspend Claude Fable 5 and Mythos 5?

Commerce acted after Amazon researchers showed a prompt that got Fable 5 to name software bugs and, in one case, write exploit code. Anthropic later said every model it tested, including Claude Haiku 4.5, Claude Opus 4.8, GPT-5.4, GPT-5.5, and Kimi K2.7, could produce the same exploit demonstration, and that the bypass did not expose unique Mythos-level offensive skills.

What Is the Difference Between Claude Fable 5 and Claude Mythos 5?

They share one underlying model. Fable 5 ships with cyber and biology classifiers. Mythos 5 lifts those safeguards for vetted cyber defenders and, in later programs, life-science labs. At the June 9 launch both cost $10 per million input tokens and $50 per million output tokens, less than half the price of Claude Mythos Preview, and Fable 5’s original classifiers triggered on average in less than 5% of sessions.

What Did Executive Order 14409 Say About Licensing AI Models?

Section 3 gave agencies 60 days to build a classified cyber benchmark and let the NSA director, with other officials, name a covered frontier model. Developers could then share that model with the government for up to 30 days before other trusted partners saw it. The same section said the order did not authorize a mandatory license, preclearance, or permit for releasing new models, including frontier models.

When Did OpenAI’s GPT-5.6 Become Available to the Public?

OpenAI opened GPT-5.6 Sol, Terra, and Luna to the public on July 9, 2026, 13 days after a limited preview for a small group of trusted partners whose names had been shared with the government. The same day it launched ChatGPT Work, an agent that combines ChatGPT and Codex on the 5.6 family.

Did the Commerce Department Lift Controls on Both Models at Once?

Yes, on June 30. Mythos 5 had already been cleared on June 26 for a named list of U.S. organizations, and that product gate survived the license withdrawal. Fable 5.1, the September successor, is generally available, while Mythos 5.1 remains inside trusted-access programs, and U.S.-only inference is offered at 1.1 times standard token prices.

Lutnick’s June 30 letter said a license was no longer required for Fable 5 or Mythos 5. It also said the administration could reimpose the controls if circumstances changed or Anthropic failed the commitments that bought the lift.