Fable and Mythos become a state secret

I had barely gotten the hang of Claude Fable 5 when it stopped existing. Three days, that's how long Fable 5 and Mythos 5¹ lasted, the two most powerful models Anthropic has ever put online. They launched on June 9, 2026² and went dark on the evening of the 12th³, on orders from the US government. An export-control directive, issued at 5:21 p.m. Eastern, told Anthropic to cut off access on national security grounds. Put simply, the most advanced AI on the market just got reclassified as sensitive equipment, and I find that worrying.

The models that lasted three days

Fable 5 and Mythos 5 share one brain, a model billed as the best around on most benchmarks, as comfortable writing code as it is helping with molecular biology research². The difference is the guardrails. Fable, open to everyone, screens sensitive requests. Mythos, limited to approved organizations, doesn't, because security professionals need the model without a muzzle. That raw power is what ended up spooking Washington.

The directive covers any foreign national, on US soil or not, down to Anthropic's own non-American employees⁴. On paper, the idea is to keep the technology from leaving the country. Except Anthropic had no way to sort its users by nationality in real time, so it shut everything off, for American customers as much as everyone else. A model deemed too dangerous for foreign hands ended up out of everyone's hands. Nice work.

And the reasoning hangs by a thread. According to Anthropic, the Commerce Department acted after a rival company claimed it had jailbroken Mythos⁵, meaning it had gotten around its protections³. Anthropic says it has only seen verbal evidence of a narrow workaround, pulled off by asking the model to analyze code. The company is complying while pushing back, calls it a likely misunderstanding, and points out that no provider can guarantee a model is airtight. In the meantime, it says it's working to restore access.

Counter-proliferation finds a new target

Export controls are Cold War vocabulary. We came up with them for fissile material, centrifuges, and missile warheads. Seeing them applied to the weights of a neural network⁶, files you can copy in seconds, tells you a lot about where we are. 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 would arm a rival.

What bugs me is that the model accused of possibly helping design a weapon is also the one that promised to speed up drug research². Frontier AI cuts both ways, by nature. And the chosen response is to lock it up instead of regulating it. Once again, political and military interests come before human ones, as if caution could only be spelled with a padlock.

A two-speed world

The incident might get sorted out in a few days. It's the logic behind it that worries me. If the most powerful models stay reserved for a handful of states, the gap between the countries that build them and the ones that watch them go by will only grow. The first group pulls further ahead, the second falls years behind. It's the same hoarding you already see among private fortunes, only between countries. Technological sovereignty stops being conference filler and becomes a question of survival.

Europe, sovereign on paper

The European Union, as it happens, has just noticed. On June 3, the Commission unveiled its technological sovereignty package⁷. There's a Chips Act 2.0 for semiconductors, a law meant to triple data center capacity, an open-source strategy, and an energy roadmap. Ursula von der Leyen wants to make Europe "an AI continent." It would sound better if it hadn't shown up after the battle.

Over 400 billion euros in total, according to the trade press⁸. Sounds huge, but the most advanced chips are still dominated by America's Nvidia and by the Asian foundries TSMC and Samsung, the factories that manufacture chips for everyone else. No plan, however big, builds the missing factories, talent, and supply chains overnight. By the analysts' own reckoning, the package nudges the long-term trajectory without touching American dominance before 2030.

So what happens to the money? For a champion like France's Mistral AI to even think about competing, it needs enough public money, money that actually gets paid out, not summit announcements. As long as it's all speeches, these companies do what they can to survive and raise their billions wherever they can, often across the Atlantic. We'll be in no position to act surprised. I'm French, I own the bias, and I'll say it anyway. A continent that can write its own rules should be able to write its own models.

One signature, on a Friday at 5:21 p.m., and the tool thousands of people were discovering that same morning vanishes from the entire planet. That's dizzying. Fable and Mythos, true to their names, got the shelf life of a bedtime story. I don't know if 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. How long can a country afford not to have its own?

Notes

1. Fable and Mythos, a fable and a myth. Anthropic had a knack for foreshadowing. It named two models after kinds of stories, and the public life of both already reads like one. ↩

2. Anthropic, "Claude Fable 5 and Claude Mythos 5", June 9, 2026. ↩ ↩² ↩³

3. Anthropic, "Statement on the US government directive to suspend access to Fable 5 and Mythos 5", June 12, 2026. ↩ ↩²

4. CNBC, "Anthropic disables access to Fable 5 and Mythos 5 to comply with government directive", June 12, 2026; see also the New York Times coverage of the same episode. ↩

5. The word comes from phones, where you'd "jailbreak" an iPhone to install whatever Apple wouldn't allow. For an AI, it means getting it to say what its guardrails refuse, usually through creative phrasing. Back when jailbreaking your phone was about changing the icons, it all seemed so innocent. ↩

6. The weights are the billions of numerical settings that, once training is done, are the model. Whoever has the file has the model, hence the anxiety. ↩

7. European Commission, "Strengthening Europe's tech sovereignty", June 3, 2026. The data center law is called the Cloud and AI Development Act. ↩

8. European Business Magazine, analysis of the EU sovereignty package, June 2026. The breakdown gives 120 billion euros for semiconductors, 200 billion for data centers by 2036, 100 billion for cloud and AI, and 2 billion for open source. ↩