Anthropic’s Mythos 5 Is Back, but the New Conditions Tell the Bigger Story

Anthropic’s most powerful AI model, Mythos 5, came back online for a small group of vetted US institutions on June 26, and Anthropic restored global access to the more public-facing Fable 5 on July 1. The three-week Mythos 5 export ban that started with a 5:21pm ET phone call from the Commerce Department on June 12 ended with a letter, a new safety classifier and a standing set of conditions for every frontier model Anthropic ships from here. Two of the three leading US frontier AI labs are now operating inside the same regulatory perimeter the Trump administration said it would never build.

Fable 5 reached Pro, Max, Team and select Enterprise plans at the restored price, with weekly usage capped at 50% through July 7. Anthropic posted on X that it had “received notice that the Department of Commerce has lifted export controls on Claude Fable 5 and Mythos 5.” Per Commerce Secretary Howard Lutnick, the deal is meant to keep American AI ahead of China. Per Anthropic, the deal also redrew the rules for every model the company releases next.

What Mythos Actually Is

Mythos is Anthropic’s most advanced series of AI models, built and based in San Francisco, and run by co-founder and chief executive Dario Amodei. The series is especially good at two things: finding weaknesses in computer systems, and fixing them. Because that skill could be repurposed by hackers, Anthropic first gave Mythos to a small group of trusted partners inside the United States, then opened it up internationally, and on June 9 released a safer public version called Claude Fable 5.

Fable 5 is built for the broader Claude user base, with safety filters meant to stop misuse of its cyber capabilities. Anthropic published model documentation describing the launch-day guardrails, the partner vetting process and the pushback from developers who thought the filters were too tight; what Fable 5’s launch-day guardrails actually block lays out the contested ground.

Mythos sits behind that public surface. Anthropic says the program, called Glasswing, ships the more powerful model to vetted cybersecurity partners – about 100 institutions in the United States as of late June – under a separate access regime. The June 9 split between Fable 5 (broad release) and Mythos 5 (vetted release) was the technical setup that made the next nine days possible.

Three Days from Launch to Lockout

Claude Fable 5 went live on June 9, 2026. Three days later, on June 12, at 5:21pm ET, the US Department of Commerce sent Anthropic an export control directive suspending both Fable 5 and Mythos 5 for any foreign national, anywhere in the world, including Anthropic’s own foreign employees. “The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance,” Anthropic wrote in Anthropic’s full statement on the export control directive.

The directive gave Anthropic no reliable way to verify nationality in real time, so the company pulled the plug on both models for everyone, US citizens included. Per the original report on the June 12 suspension, Commerce did not name the national security concern in writing; the only evidence Anthropic received was verbal.

Anthropic publicly disagreed with the move, calling the underlying jailbreak finding narrow and saying other publicly available models could produce the same result. “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 wrote. The company later tested the technique itself and said the same code path was reproducible in Claude Opus 4.8, GPT-5.5 and Kimi K2.7. Per Fable 5’s restoration timeline and the new classifier behind it, the finding was “a borderline case for Fable 5’s safeguards” rather than a Mythos-level capability.

The UK government’s AI Security Institute ran an independent test and found Mythos could exploit defenses and systems 73% of the time. The BBC reported that figure from Gina Neff, professor of responsible AI at Queen Mary University London, who called the result “a step change in capability in cyber security.”

1. June 9, 2026 – Anthropic launches Claude Fable 5 publicly; Mythos 5 to trusted partners under Glasswing.
2. June 12, 2026 at 5:21pm ET – Commerce directive suspends both models for any foreign national, anywhere.
3. June 26, 2026 – Mythos 5 cleared for 100+ US institutions under Glasswing.
4. June 30, 2026 – Commerce withdraws the June 12 controls on both models.
5. July 1, 2026 – Fable 5 returns to global users under new conditions.

Why Anthropic Fought Back

The June 12 directive did not land in a vacuum. Per the Economic Times, Anthropic had ties to Democrats and pushes a pro-safety stance on AI. In February, the company refused the Pentagon’s demand to let the Defense Department use its AI without limits on surveillance and weapons. President Donald Trump responded by ordering every federal agency to stop using Anthropic’s products and calling the company “Leftwing nut jobs” on social media. Defense Secretary Pete Hegseth labelled Anthropic a “supply chain risk,” the first time a US company had received such a designation publicly. The full account of that earlier refusal sits in Anthropic’s prior refusal of the Pentagon’s terms.

Some observers believe the Fable 5 and Mythos 5 curb was politically motivated and connected to that earlier spat, not to AI safety. Anthropic itself framed the move as a misunderstanding. According to how Amazon’s jailbreak finding reached the Commerce Department, the underlying information came from an Amazon researcher and was passed to the government via CEO Andy Jassy. Anthropic remains locked in a separate lawsuit against the Pentagon over the supply-chain-risk label, and a US judge has ruled the directive cannot be enforced while the case continues.

Anthropic’s posture going in was clear: even its own foreign employees were covered by the directive. The dispute ended with conditions, not a clean win on the merits. Mythos 5’s interim approval for 100+ US institutions captures the halfway point the company passed through on the way back.

• 73% – UK AISI’s measured cyber-exploit success rate for Mythos.
• 99% – Anthropic’s new safety classifier blocks the specific jailbreak technique.
• 30 days – voluntary pre-release review window under Trump’s June 2, 2026 executive order.
• 100+ – US institutions cleared to use Mythos 5 on June 26 under Glasswing.

The Conditions Now Attached to Every Frontier Model

Per the Commerce letter that ended the three-week blackout, Lutnick’s letter framed the lift as joint work. Anthropic agreed to detect and address security risks on its own models, collaborate with the US government on protocols and standards for Mythos, Fable and future models, and alert the government to any malicious activity. Anthropic said it would “scale up our government collaboration” and “stand up dedicated Anthropic teams to work on shared government priorities.”

A new safety classifier, trained by Anthropic and tested by the Commerce Department’s Center for AI Standards and Innovation, blocks the original jailbreak technique in over 99% of cases. Requests that would trigger the technique now route to Claude Opus 4.8 with a block notice shown to the user. CAISI told Anthropic the safeguards are “extraordinarily strong.” Anthropic acknowledged a trade-off: the new classifier “comes at the cost of flagging benign requests more often during routine coding and debugging tasks.”

Anthropic said the agency had given the company “verbal evidence of a potential narrow, non-universal jailbreak” before the June 12 directive, and argued the same evidence should not have been enough to pull a commercial model. The fix the company shipped pairs the new classifier with a broader push: Anthropic is working with Amazon, Microsoft, Google and other Glasswing partners on a consensus framework for scoring jailbreak severity, a shared yardstick for what counts as a national-security incident. Dean Ball, a former White House AI adviser and soon-to-be OpenAI employee, said the underlying executive order that asks AI companies to voluntarily submit frontier models for up to 30 days of pre-release review had “created a de facto involuntary licensing regime for frontier AI.”

Anthropic has agreed to proactively detect and address security risks associated with the models; to work diligently with the US government on protocols and standards and releases for Mythos, Fable, and future models; and to inform the US government of any malicious activity.

The blockquote above is taken from Lutnick’s letter to Anthropic, as reported by the Guardian.

OpenAI in the Same Net

One day before Anthropic’s announcement, OpenAI launched its next-generation model family, GPT-5.6 Sol, Terra and Luna, and disclosed the Trump administration had asked the company to limit initial access to “a small group of trusted partners.” OpenAI framed it as a temporary compromise, writing that it would take “this short-term step because we believe it is the strongest path to broader availability in the coming weeks.” Chief executive Sam Altman pushed back harder on X, posting that the arrangement “isn’t quite the process that we think is optimal” and that OpenAI would work with the government “to attempt to get to a transparent, reliable process for early access.” Altman told the Guardian that extensive safety testing “is not a bad idea. I just don’t like the idea of the government picking the customers.” OpenAI’s GPT-5.6 pricing sets Sol at $5 per million input tokens and $30 per million output tokens, Terra at half that, and Luna at $1 and $6.

OpenAI has not named the trusted partners, and Anthropic’s restrictions are tighter for international users than for its own US customer base. Two of the three leading US frontier labs now sit inside a regime that did not exist in this form a month ago, and both have signed letters that put them inside it.

The Knock-On Effect: China Catches Up

While Washington and Anthropic were negotiating, Z.ai, the Beijing lab formerly known as Zhipu AI, released GLM-5.2 on June 13, 2026, under an MIT license. CNBC reported that GLM-5.2 “can perform on par with top US labs on some cyber benchmarks, even equaling Mythos’ capabilities” within days of the US restrictions landing.

Marc Andreessen, a White House adviser and a16z co-founder, posted on X over the weekend of June 27: “Many smart people/AI insiders are saying GLM-5.2 is the first Chinese AI model to match and often beat the American big lab public AI models with no compromises. Incredible timing given current events.”

Even David Sacks, the former White House AI and Crypto Czar who had publicly feuded with Anthropic over safety, weighed in. Per CNBC, Sacks posted above a Wall Street Journal headline stating that China had matched Anthropic in cybersecurity: “A year ago, President Trump declared that America was in a global AI race and that the way to win it was to be pro-innovation, pro-infrastructure, pro-energy, and pro-export. President Trump was exactly right; we deviate from that strategy at our peril.” Coinbase CEO Brian Armstrong wrote on X that his company is using open-weight models such as GLM-5.2 and Kimi K2.7, cutting nearly half its AI spending despite greater token use. Flo Crivello, CEO of AI startup Lindy, told CNBC he switched his company off Anthropic’s Claude models and moved 100% of its traffic to DeepSeek.

Five Eyes intelligence agencies in Australia, the US, the UK, New Zealand and Canada warned in late June that frontier AI “accelerates the speed, scale, and sophistication of cyber threats.” Their assessment: “Frontier AI models are anticipated to exceed current industry expectations, fundamentally transforming both offensive and defensive cyber capabilities. The timeline is not years, it is months.” how GLM-5.2 closed the gap on Mythos during the standoff covers the timing in detail, and GLM-5.2 topping Mythos on IDOR detection at one-sixth the cost walks through the vulnerability benchmarks.

• GLM-5.2 (Z.ai, MIT-licensed open-weight) – matches Mythos on cyber benchmarks per CNBC.
• Kimi K2.7 (Moonshoot AI) – used by Coinbase alongside GLM-5.2.
• DeepSeek – Lindy’s CEO moved his company’s Claude traffic to DeepSeek in June.
• Alibaba Qwen 3 – cited by Shopify and Airbnb for scaling AI features.

What Could Change in August

On June 2, 2026, Trump signed Executive Order 14409, “Promoting Advanced Artificial Intelligence Innovation and Security,” asking AI companies to voluntarily submit frontier models for up to 30 days of pre-release review. In August, the government is due to publish the criteria for which models will fall under the new security restrictions. Per the Economic Times, the August framework will settle whether Mythos-class models stay gated the way they are now, or whether the perimeter widens.

The European Commission, which gained access to Mythos earlier in June after weeks of talks, called the situation “Europe’s need for technological sovereignty.” Thomas Regnier, a Commission spokesman, told the BBC: “We take note of Anthropic’s statement and are assessing.” The international side of Mythos 5, including the European Commission’s access, remains undefined. Anthropic has called for shared standards to assess critical vulnerabilities in advanced models and said it will work with Amazon, Microsoft, Google and others on that effort; the company says the same yardstick should apply to Chinese and American frontier models alike.

Frequently Asked Questions

When did Fable 5 come back online?

Anthropic restored global access to Claude Fable 5 on July 1, 2026. Through July 7, the rollout was capped at 50% of weekly usage limits on Pro, Max, Team and select Enterprise plans, after which Fable 5 moves to standard usage credits. Mythos 5 had already returned in limited form on June 26, when Commerce cleared it for about 100 US institutions under the Glasswing program.

Why did the US suspend access to Fable 5 and Mythos 5?

The Commerce Department issued an export control directive on June 12 at 5:21pm ET citing a jailbreak that bypassed Fable 5’s safeguards. The underlying finding came from Amazon researchers, passed to the government via Amazon CEO Andy Jassy. Anthropic said in its public response that the vulnerability was narrow, that other models could reproduce the same code, and that “the finding of a narrow potential jailbreak” should not be cause to pull a deployed commercial model.

Was the directive about the jailbreak or about Anthropic’s politics?

It is contested. Anthropic publicly disputed the jailbreak as a narrow issue other models could replicate. Per the Economic Times, some observers believe the curb was politically motivated and connected to Anthropic’s February refusal of the Pentagon’s surveillance and weapons terms, a position Anthropic’s critics inside the administration have not endorsed on the record.

Does the same rule now apply to OpenAI’s GPT-5.6?

OpenAI launched GPT-5.6 Sol, Terra and Luna at the Trump administration’s request, with initial access limited to a small group of trusted partners. CEO Sam Altman has said on X that the company will work with the government “to attempt to get to a transparent, reliable process for early access,” and that “I just don’t like the idea of the government picking the customers.” Both Anthropic and OpenAI are now under US government review of frontier model releases before general availability.

What is Mythos 5 still restricted from?

International release. The June 26 partial approval cleared Mythos 5 for about 100 US institutions under the Glasswing program, including major companies and federal agencies. Broader domestic and international access is still being negotiated under the same program, and the August criteria from the executive order may reset who else gets access.