Fable and Mythos, the day AI got classified

Three days. That's how long Claude Fable 5 and Claude Mythos 5^[1] lasted, the two most powerful models Anthropic has ever put online. They launched on June 9, 2026^[2] and went dark on the evening of the 12th^[3]. The cause was a US government export-control directive, issued at 5:21 p.m. Eastern, that ordered Anthropic to suspend access on national security grounds. By the time most of us had found our footing with the most advanced AI on the market, it had already been reclassified as sensitive equipment.

The models that lasted three days

Start with what Anthropic had just shipped. Fable 5 and Mythos 5 share one brain, a model from the new Mythos class that the company billed as state of the art on most benchmarks, as comfortable writing code as it is assisting molecular biology research^[2:1]. Fable, open to everyone, keeps guardrails that screen sensitive requests; Mythos, limited to approved organizations, drops them, meant for the security professionals who need the model without a muzzle. That raw capability is what eventually spooked Washington.

The directive targets a narrow group: any foreign national, whether or not they're on US soil, down to Anthropic's own non-American staff^[4]. On paper, the point is to keep the technology from leaving the country. In practice, Anthropic couldn't sort its users by nationality in real time, so it shut everything down, for American customers as much as for everyone else. A model deemed too dangerous for foreign hands ended up out of everyone's hands, Americans included.

The reasoning is thin. By Anthropic's account, the Commerce Department acted after a rival company claimed it had jailbroken Mythos^[3:1]. Anthropic says it has seen only verbal evidence of a narrow, non-universal jailbreak, produced by asking the model to analyze code. The company is complying without fighting the order, calls it a likely misunderstanding, and points out that no provider can promise perfect resistance to that kind of exploit, all while saying it's working to restore access.

Counter-proliferation finds a new target

Export controls are Cold War vocabulary. We built them for fissile material, centrifuges, and missile guidance, the kinds of things whose movement is measured in strategic balances. Watching them apply to neural network weights, files you can copy in seconds, says a lot about the shift underway. After space and quantum computing, AI joins the list of technologies a great power keeps to itself, on the theory that letting one copy out the door would arm a rival.

The discomfort starts here. The same model accused of helping design a weapon is the one that promised to speed up drug research and molecule discovery^[2:2]. Frontier AI is dual-use by nature, and the response on the table is to lock it away rather than govern it. Once again, political and military interests won out over human ones, as if caution could only be spelled with a lock.

A two-speed world

The incident itself may be sorted out within days. What unsettles me is the logic it sets in motion. If the most powerful models stay the preserve of a few states, the gap will widen between the countries that build them and the ones that watch them go by. The first group pulls further ahead, the second loses years, and you get, transposed to nations, the same hoarding you already see among private fortunes. Technological sovereignty, national or carried by a union of countries, stops being conference-panel filler and becomes a question of survival.

Europe, sovereign on paper

The European Union, as it happens, just woke up to this. On June 3, the Commission unveiled its technological sovereignty package: a Chips Act 2.0 for semiconductors, a Cloud and AI Development Act meant to triple data center capacity, an open-source strategy, and an energy roadmap tied to sovereign AI models^[5]. Ursula von der Leyen wants to make Europe "an AI continent," a line that would land better if it hadn't arrived after the battle.

The figures floated in the trade press show the size of the hill to climb: 120 billion euros for semiconductors, 200 billion for data centers by 2036, 100 billion for cloud and AI, 2 billion for open source^[6]. The trouble is that the most advanced chips are still dominated by America's Nvidia and Asia's foundries, TSMC and Samsung, and no plan, however large, builds the factories, the talent, and the supply chains overnight. By the analysts' own reckoning, the package shifts the long-term trajectory without touching American dominance before 2030.

That leaves the question of what you do with the money. Backing a champion like France's Mistral AI takes more than announcements and summits. It means putting in enough public money, actually disbursed, for the company to even think about competing. As long as it stays at the level of speeches, these firms do what they can to survive and go raise their billions wherever they can find them, often across the Atlantic, and we'll have little standing to act surprised afterward. I'll own the bias here, I'm French, and I hold to the idea that a continent able to write its own rules should also be able to write its own models.

There's something dizzying about a signature dropped on a Friday at 5:21 p.m. wiping out, across the entire planet, the tool thousands of people had discovered that same morning. Fable and Mythos, true to their names, got the shelf life of a bedtime story. I don't know whether access comes back next week or never, but the next breakthrough will meet the same fate, and the countries without a model of their own will watch it go dark with no say in the matter. Which leaves one question, and it isn't a technical one: how long can a country afford not to have its own? So, how does it sit with you, an AI switched off with the stroke of a pen?