When an AI Can Vanish in Ninety Minutes
You have spent months fine-tuning your setup, choosing your tools, building automations you rely on every day. Then one morning, one of those tools simply vanished. Not an outage, not a bankruptcy, not a bug. A political decision.
That is exactly what happened in June 2026 to Anthropic's two most powerful models, Fable 5 and Mythos 5. Cut off for the entire planet, the United States included, in under ninety minutes following a government directive. These models had been available for only four days.
What this episode reveals goes far beyond two models being taken offline. Since June 2026, the release of the most advanced AI is no longer decided by the labs alone. It is filtered upstream by the US government. This is the arrival of government-controlled AI, and it changes a great deal.
As I write these lines, in early July 2026, here is what we are going to break down together: the timeline of this shift, the reason Washington gave, what it means concretely for anyone who cares about controlling their own tools, and above all who really benefits from this lock. I will tell you right away, the answer has a name: China.
The Month That Upended AI: The Timeline
June 12: Anthropic Unplugged Almost Instantly
The US Department of Commerce orders Anthropic to cut off access to Mythos 5 and Fable 5 for any foreign national, company employees included, in the name of national security. Anthropic does not do things by halves and pulls both models outright, including for American users.
This is a first. Never before had export control authority, a tool built for physical goods and weapons, been applied to a commercially deployed AI API. The concrete result for customers: models that evaporated overnight, and refunds to go with it.
The Trump Executive Order: Thirty Days of Review Before Any Release
In the background, an executive order signed by Donald Trump early in the month. It requires labs to submit their most advanced models to a government review thirty days before any release. The catch is that the terms of this review remain undefined: we do not know who evaluates, by what criteria, or who gets the final say.
June 26: GPT-5.6 Ships Already Locked Down
Two weeks later, OpenAI launches GPT-5.6, offered in three versions named Sol, Terra, and Luna. Except that the launch is restricted access, reserved for a small circle of government-approved organizations, at the request of two White House offices: the Office of the National Cyber Director and the Office of Science and Technology Policy. No public waitlist, no open sign-up.
Sam Altman, OpenAI's CEO, reportedly told his teams, according to Tech Times, that access would be approved customer by customer during this phase. Notably, OpenAI stated in black and white that it does not want this kind of process to become the norm, stressing that it deprives developers, companies, and cyber defenders of tools they need. Here we have a rare admission: a lab enforcing a constraint while publicly saying it should not exist.
June 30 and July 1: The Lock Loosens, Fable 5 Returns
A dramatic turn at the very end of the month. On June 30, the Department of Commerce lifts the export controls on both models. Right after, Anthropic announces the worldwide return of Fable 5 as of July 1, across all its platforms. The consumer version becomes accessible again, but not under the same conditions as before.
The model returns fenced in by new safeguards. A reinforced safety classifier now blocks certain cybersecurity-related requests, to the point of automatically routing routine tasks like coding or debugging over to Opus 4.8. Anthropic owns the trade-off: this stricter filter, out of an abundance of caution, also turns away more perfectly legitimate requests while it is being fine-tuned.
Careful not to mix everything up, though. Mythos 5, the far beefier big brother, has not come back for everyone. It remains reserved for a set of authorized American organizations focused on critical infrastructure and cyber defense. One detail that stings: during the blackout, the NSA itself had lost its access to Mythos, according to the specialized outlet Nextgov, even though its analysts were using it to hunt for flaws in classified systems.
Washington's Argument: Cybersecurity, and Its Limits
Before we cry foul, let us take two minutes to understand the logic. It is not completely crazy.
These models are not toys. GPT-5.6 Sol passed 96.7% of OpenAI's internal offensive tests, the exercises where an AI is asked to carry out real cyberattacks from the command line. All three versions in the lineup reach the high cyber risk level defined by OpenAI, a first for models this accessible. Mythos and Fable, for their part, were considered the highest-performing models on the market.
The worry is about dual use. A model able to find and chain vulnerabilities gives a real boost to an attacker. Seen through the lens of national security, wanting a head start before the entire world, hostile actors included, gets its hands on it, that is defensible.
Where the shoe pinches is the method. A mechanism designed for goods and weapons is being applied to software served in the cloud, with no dedicated legislative basis, no appeals process, with minimal notice. Neil Chilson, former chief technologist at the FTC, sums up the problem with an image: permanently holding a sword of Damocles over every lab will mostly push them to slow down their releases, depriving the public of useful tools. And let us remember the block did not only hinder foreign actors. It hindered the NSA.
The Real Lesson: A Model You Cannot Host Is a Model That Can Be Taken From You
Here is the point that should give us all pause, and it speaks directly to the DNA of this site.
The difference between a closed model and an open model had until now been mostly philosophical. It has become an architecture decision. A closed model like Opus, Mythos, or GPT lives entirely on the vendor's servers, accessible through an API that the company, or a regulator, can cut off with a single gesture. An open source model, or more precisely an open weights model, publishes its parameters under license: you download them and run it on your own hardware, in your own environment.
A model you cannot host is a model that can be taken from you.
The smart home community has held this reasoning for years. It is exactly the debate between the connected device that depends on a distant server and the local Zigbee or Matter sensor that keeps working even when the internet goes down. The AI world has just learned, the hard way, the lesson Home Assistant users internalized long ago: what you do not control, someone else controls, and they can pull the plug.
Building a critical automation, or an entire part of your business, on a remote AI you do not control means creating a single point of failure beyond your reach. This holds for a company as much as for your connected home if you route every command through an assistant hosted elsewhere.
The Big Winner Has a Name: China
While the United States locks things down, one player is moving its pieces forward, and it is not necessarily where you would expect.
In mid-June, the Chinese lab Zhipu (operating under the Z.ai brand) released GLM-5.2: open weights, MIT license, available for download on Hugging Face. The model sits about one point behind Anthropic's Opus 4.8 on a key agentic benchmark, for roughly a fifth of the cost, according to a CNBC analysis. Downloadable, self-hostable, and sold with an argument that lands right on cue: no regional limits.
The timing is almost comic. GLM-5.2 arrived in nearly the same days that Anthropic was going dark. Zhipu even wrapped its release in a crystal-clear message: cutting-edge intelligence should not be removable at any moment by a few rules. Hard to craft better publicity at the expense of the American lockdown.
And this is the whole strategic paradox. By turning its best models into tools a Commerce Secretary can switch off by letter, even temporarily, the United States has turned itself into a supply risk. A model no one can unplug becomes a commercial argument. Companies planning over several years are starting to see the non-revocable option as the safest choice, regardless of who tops the rankings at any given moment. Traffic to GLM-5.2 on developer platforms is in fact climbing faster than after the DeepSeek episode a year ago.
Let us be fully honest, because this is no silver bullet. Self-hosting GLM-5.2 has a real cost in hardware and engineering, these models are enormous. Going through Zhipu's cloud API reintroduces a dependency, this time under Chinese law. And open weights does not mean fully open code: the parameters are published, not necessarily the entirety of the training data. But on the criterion that is becoming central, keeping control, the gap with a closed model is gaping.
What All This Signals: AI Becomes a Matter of State
This June did not just bring two models to a halt, it enacted a change in nature. Until now, an AI model was a product, a service you subscribe to. By drawing the weapon of export controls, the one usually reserved for arms and sensitive components, Washington has just filed cutting-edge AI into an entirely different category: that of a state's strategic assets. On the same level as oil, semiconductors, or a military technology.
And once that line is crossed, there is no going back. The proof: OpenAI did not even wait to have its hand forced. The throttled launch of GPT-5.6 is the voluntary, preventive version of the same mechanism. Anthropic, for its part, states in black and white that it is deepening its cooperation with the government, with access to its models before release and jointly conducted testing. Translation: the state is no longer a regulator watching from afar, it is settling in for the long haul, inside the loop. To my mind, the next releases of frontier models will look less and less like a product launch and more and more like an arms deal, with prior review and administrative green light.
The first concrete milestone is coming fast. The US government has set itself a deadline in August 2026 to build a real evaluation framework for models with advanced cyber capabilities. That is where everything is decided: either this process becomes a clear, predictable procedure, or it stays what it is today, a case-by-case negotiation, lab by lab, letter after letter. The most likely scenario, in all honesty, is a wobbly in-between for quite a while.
The geopolitical consequence is already before our eyes: the AI world is fracturing into two blocs. On one side, a Western technological frontier, brilliant but locked, whose access depends on an administration's goodwill. On the other, a Chinese ecosystem, open and borderless, a notch behind on paper but catching up at full speed, and that no one can unplug. Zhipu is not advancing alone, either: Kimi, MiniMax, Qwen, and DeepSeek are pushing in the same direction, and Zhipu's boss is already announcing an open model at the Fable or Mythos level before the end of the year. Every week of American lockdown is a week of free publicity for them.
One big absentee remains in this duel: Europe. During the block, it was foreign users who were targeted first. From Paris, Berlin, or Amsterdam, you depended for your best AI tools on decisions made in Washington, over which you had absolutely no say. This is no longer a theoretical worry. Austria has in fact already written to the European Commission to urge it to bring about a framework capable of welcoming Anthropic onto the continent, according to Bloomberg and Reuters. The phrase digital sovereignty, long waved around in a vacuum in political speeches, has just taken on a very concrete meaning for millions of users and businesses that decided nothing, and yet bore the brunt.
Conclusion: We Have Seen Nothing Yet
The cyber risk argument is not stupid, these models can now dig up and exploit flaws almost on their own. But the method used, opaque, with no recourse, with a weapon meant for goods and armaments drawn against a consumer API, to the point of cutting off the NSA itself and every foreign user in a single gesture, misses its target entirely. It protects hardly anyone. Above all it pushes the rest of the planet toward open models, and therefore, right now, toward China.
The real lesson goes far beyond this episode. In a few weeks, cutting-edge AI has officially left the products aisle to join that of national sovereignty stakes. The state has entered the loop, and it will not leave. What has just played out with Fable and Mythos is not an isolated accident: this placing of AI under state control is only the first chapter of a story that will occupy the entire coming decade.
Two or three years from now, the real question will no longer be which model is the smartest, but which one you are still allowed to use, from where, and under what conditions. We are entering an era where an AI's raw power may matter less than where you hold the tap. And for now, that tap is very far from Europe.
How far are you willing to entrust your uses, personal and professional, to an AI that someone else can switch off overnight?
