Introduction

On June 12, 2026, the Trump administration invoked export control authority to force Anthropic offline. The process took ninety minutes. The next day, Beijing-based Zhipu AI shipped GLM-5.2 under an MIT license — no restrictions, free to download, free to run on any hardware, anywhere on earth. Within a week, GLM-5.2 had climbed to the top of the openly accessible model leaderboards, Zhipu's market capitalisation had crossed HK$1 trillion, and the most capable model legally accessible to the majority of the world's developers was a free download from a company that sits on Washington's own trade blacklist. On June 30, 2026, CNBC reported that the Trump administration's crackdown on Anthropic's leading AI models was looking like a gift to the country's top adversary in the AI race: China. The same day, the Commerce Department fully lifted the export controls it had imposed eighteen days earlier. The controls were gone, but the damage they caused was not.

That is not an argument for ignoring AI security risks. It is an argument for the proposition that the manner in which the US government executed its response to a narrow jailbreak vulnerability caused strategic harm that dwarfed any security benefit it plausibly achieved — and that understanding exactly how that harm accumulated matters for anyone trying to build a US AI policy that works rather than one that merely performs decisiveness.

What Happened During the Eighteen Days

The Fable 5 affair, from first shutdown order to full restoration, lasted eighteen days. The sequence of events within those eighteen days is the most accurate measure of what the export control action actually cost.

On June 12, Anthropic disabled Fable 5 and Mythos 5 for every user on earth, including its own non-citizen employees, because the Commerce Department's directive barred access for any foreign national and the company had no way to verify citizenship across its global customer base in real time. The trigger was a narrow jailbreak vulnerability flagged by Amazon researchers to Treasury Secretary Scott Bessent — a technique for extracting information about cyberattacks that Anthropic described as producing behaviour also obtainable from other public models, including OpenAI's GPT-5.5. The government did not disclose the specific vulnerability in writing. It issued an order and set a deadline.

On June 13, Zhipu released GLM-5.2. The timing was not accidental. The company had been preparing the release in advance, but its decision to frame the launch explicitly as an unrestricted alternative at the precise moment when the most capable US model had gone dark was a market signal the global developer community received immediately. On Arena.ai's Code Arena leaderboard, GLM-5.2 climbed to the second-highest position overall — and to the top of openly accessible models, because Fable 5 had been removed from the ranking following the ban. Zhipu's stock surged 48% intraday on June 15. JPMorgan raised its price target from HK$950 to HK$1,400 and named the company an AI winner. Bank of America initiated coverage with a buy rating. By June 15, Zhipu's market capitalisation had crossed $128 billion.

On June 22, the Council on Foreign Relations published an essay by Matthew Ferren, a former Department of Defense official who developed strategies and plans for employing military cyberspace operations. His title was direct: "The U.S. Is Losing the AI Credibility War — to Itself." His core argument was that the administration had undermined the proposition that American AI could be trusted by allied governments and foreign enterprises, at the precise moment when that trust was the United States' most valuable and hardest-to-replicate competitive asset. Ferren was not alone. Sam Bresnick, a research fellow at Georgetown's Center for Security and Emerging Technology, called the developments "a pretty good wake-up call." Saif Khan, who helped write earlier US export rules, said curbing sales while China builds its own tools rewards Beijing and weakens US cyber defenses. The assessment from multiple former officials was consistent: the action was strategic self-harm.

The Enterprises That Moved and Did Not Come Back

The geopolitical abstraction of the Fable 5 affair becomes concrete when you examine what individual enterprises actually did when the models went dark.

Flo Crivello, CEO of AI startup Lindy, did not wait for political resolution. He switched his company off Anthropic's Claude models entirely during the ban and moved 100% of its traffic to DeepSeek. "We did it, and you could see that cost curve go down, like, crash to the ground," he told CNBC. The cost reduction was immediate and dramatic. Lindy has not signalled an intention to reverse that decision simply because Fable 5 has been restored. Coinbase CEO Brian Armstrong wrote publicly on X that his company was utilising open-weight models including GLM-5.2 and Kimi K2.7, allowing it to cut nearly half its AI spending despite greater token use. Shopify and Airbnb had already touted the benefits of Alibaba's Qwen 3 for scaling AI features before the Fable 5 ban; the ban accelerated a trend that was already forming.

Major companies routing their AI traffic through Chinese open-weight models is not merely a commercial development. It is a supply chain shift with national security implications that cut in the opposite direction from what the administration's action intended. Travis Lanham, co-founder of AI security startup Armadin, told CNBC his company was experimenting with GLM-5.2 and Kimi K2.7 from Moonshot AI and that the models were showing improved capabilities for cybersecurity use cases including analysing reconnaissance data and creating exploit code. The action that was justified in the name of preventing Chinese access to US cybersecurity AI capability drove US companies toward Chinese cybersecurity AI models in the open-weight market where no export control mechanism applies.

The open-weight architecture is the piece of this that the administration's export control framework cannot reach. As Travis Lanham put it simply: "With the open-weight models, it's kind of the Wild West." A company that downloads GLM-5.2's weights under its MIT licence and runs the model on its own servers has no vendor sitting between it and the model. There is no API to shut down. There is no foreign national access to block. The control regime the US government constructed around frontier cyber-AI assumes a vendor sits between the model and the user. Open weights eliminate the vendor. Restricting a closed American model while a comparable Chinese open-weight alternative is freely downloadable does not reduce the aggregate cybersecurity risk in the world. It shifts the provider.

What GLM-5.2 Actually Achieved in the Window

The technical claims about GLM-5.2's capabilities during the Fable 5 gap deserve honest assessment, because both overclaiming and underclaiming those capabilities misrepresents the nature of the strategic problem.

GLM-5.2 is a 744-billion-parameter mixture-of-experts model, trained entirely on Huawei Ascend chips under a process that included no Nvidia hardware — a direct demonstration that Washington's export controls on advanced semiconductors have not prevented the development of competitive frontier models in China. On the Artificial Analysis Intelligence Index v4.1, GLM-5.2 scored 51 points, placing it ahead of MiniMax-M3 and DeepSeek V4 Pro, and ahead of Google's Gemini 3.1 Pro Preview. On Design Arena, it took first place. On the Agent Arena, it ranked first among open models by a wide margin. On SWE-bench Pro, it scored 62.1, seven points behind Claude Opus 4.8 at 69.2. On broader reasoning tasks, it still trails the US frontier. On the specific task of finding software vulnerabilities, The Wall Street Journal reported that researchers found GLM-5.2 even with Mythos at scanning code for vulnerabilities in some scenarios.

Marc Andreessen wrote on X over the weekend of the ban: "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." Jefferies strategist Christopher Wood wrote in a report to clients that GLM-5.2 "is almost equal to Anthropic as a competitor for the corporate market and is just one quarter of the cost in terms of token." Elon Musk predicted on X that Chinese labs would reach Fable-level capability by the first quarter of 2027, responding to a user's question. Zhipu founder Jie Tang replied directly to Musk: "Won't take that long."

The CFR's assessment, published by former DoD official Matthew Ferren, drew the strategic conclusion plainly. The government's action did not eliminate any cybersecurity capability from the global market. GLM-5.2, built by a company on Washington's trade blacklist, trained on Huawei chips, and released under a permissive open licence, is now the most capable freely downloadable model in the world. The capability the administration sought to contain is available to anyone with a GPU cluster and the willingness to run it. What the administration succeeded in containing was the American company's access to its own customers.

The Credibility Cost That No Restoration Fully Repairs

The export controls were lifted in full on June 30, eighteen days after they were imposed. Anthropic confirmed the restoration on X: "We're grateful to our users for their patience, and to everyone who worked with us on redeploying the models." Commerce Secretary Lutnick's June 26 letter, which began the restoration process, noted that Anthropic had committed to work with the government on protocols and standards for future model releases. That commitment is the lasting institutional change the affair produced — a requirement for Anthropic to notify and coordinate with the government before major model releases.

Whether that coordination requirement produces better national security outcomes than the alternative depends entirely on whether the government can build and maintain the technical expertise to evaluate model risks rapidly and accurately, produce specific written findings rather than verbal complaints, and exercise its authority through a defined and transparent process rather than ninety-minute phone calls. None of those prerequisites were in place before June 12. Whether they are in place now is a question the next model release will answer.

What is already clear is that the credibility cost of the affair will not be fully repaired by the restoration announcement. Lizzi Lee, a fellow on the Chinese economy at the Asia Society Policy Institute's Centre for China Analysis, was direct in her assessment to the South China Morning Post: "I think the Zhipu rally tells you the market immediately understood the opening that Anthropic just created." The opening was created by the US government's action, not by Anthropic. And the market's response — enterprises diversifying onto Chinese open-weight models, capital flowing to Zhipu, French President Macron warning that no country would purchase American AI if it could be switched off at any moment, Canadian Prime Minister Mark Carney describing the episode as a lesson in the danger of over-reliance on any single nation's technology — reflected a judgement about the reliability of American AI as infrastructure that a restoration announcement cannot simply undo.

By May 2026, before the Fable 5 affair, Chinese open-weight models already accounted for 61% of all tokens consumed on OpenRouter, the largest neutral large language model router. Four of the five most-used models on the platform were already Chinese. The ban did not cause that shift. It accelerated it, validated it, and removed the friction that might have slowed some enterprises from committing to Chinese open-weight alternatives they had already been evaluating.

What a Coherent AI Security Policy Actually Looks Like

The honest lesson from the Fable 5 affair is not that the government should leave frontier AI unregulated. It is that the government's current approach — opaque emergency authority exercised at ninety-minute notice without detailed written findings, justified by a narrow jailbreak vulnerability that Anthropic said also existed in other publicly available models — is not a security policy. It is an emergency measure that can be triggered by political dynamics as easily as by genuine technical risk, with consequences that the government's own national security establishment has described as counterproductive.

A coherent policy has several properties that the current framework lacks. It defines in advance what categories of vulnerability are sufficient to justify restricting a commercial AI product, so that companies can design toward those standards rather than discovering after launch that the product fails an unspecified test. It requires the government to produce specific written technical findings — ideally within 24 hours of identifying a concern — so that the company has a genuine opportunity to assess whether the concern is accurate and whether a technical fix is feasible. It establishes a fast-track legal review mechanism so that a company that believes the order rests on a misunderstanding, as Anthropic said of the June 12 directive, can seek judicial review on a timescale measured in days rather than months.

It also, critically, is designed with the open-weight reality in mind. A policy that restricts the best-governed American closed model while leaving comparably capable Chinese open-weight models freely downloadable under MIT licences does not protect American cybersecurity infrastructure. It transfers a portion of that infrastructure's AI supply chain to entities operating outside any governance framework the US government controls. Saif Khan, who helped write earlier export rules, said the current approach rewards Beijing and weakens US cyber defenses. He is right. A security policy that produces that outcome has its strategic logic backwards, regardless of how decisive the individual action looks in the ninety minutes during which it is enforced.

Conclusion

The Fable 5 affair cost America eighteen days of access to its most capable AI models for every non-US-citizen developer, researcher, enterprise, and government on earth. During those eighteen days, the most capable freely available model in the world became a free download from a company on Washington's trade blacklist, trained on hardware Washington tried to deny it, released under a licence Washington cannot restrict. Enterprises moved to Chinese models and reported cost reductions so significant they are unlikely to reverse them simply because Anthropic's models are available again. The credibility of American AI as reliable infrastructure — the proposition that allied governments and global enterprises can build on US AI without political disruption — took damage that restoration announcements cannot fully repair.

The technology America used to justify the action — a narrow jailbreak vulnerability that existed in other models, described by Anthropic as resting on a misunderstanding — turned out not to justify the strategic cost of exercising the authority as it was exercised. That does not mean the authority should not exist. It means the authority should be used with a level of precision, transparency, and technical grounding that the June 12 directive conspicuously lacked. China did not close the AI gap in eighteen days. But the enterprises and governments that spent those eighteen days without Fable 5 and found Chinese alternatives that cost a quarter as much per token learned something about supply chain diversification that they will not forget. That is the real cost of the Fable 5 affair, and no amount of partial credits for the government eventually getting the models restored should obscure it.