On Tuesday, Anthropic put its most capable model yet in front of the public. By Friday evening it was gone. Not because it broke, but because a government said so.
On June 12, the US Commerce Department issued an export control directive, in a letter from Commerce Secretary Howard Lutnick, ordering Anthropic to suspend Fable 5 and Mythos 5, its two most capable models, for any foreign national, inside or outside the United States. To comply, Anthropic switched both off for every customer worldwide. Its other models stayed up. Three days later the models are still dark, and Anthropic has sent senior staff to Washington to negotiate their return.
Anthropic disagrees with the order and is contesting it in public. They may well win. This post is not about whether they are right. It is about a structural fact the episode made impossible to ignore.
The switch was always there
If your AI provider sits under US jurisdiction, a US government directive can remove your access. Not your data this time. Your access. The model you built on can be switched off by a letter you never see, for reasons you are not told, with almost no notice. The customer gets no vote and no veto.
According to reporting on the order, it appears to be the first time an AI company has taken a publicly deployed model offline because the federal government told it to. The stated trigger was thin: Anthropic says the alleged jailbreak is narrow, already known, and that comparable capability is available from other public models, including OpenAI's GPT-5.5. The point here is not the merits. It is that the switch exists, and it sits in one capital.
Two failure modes, one root
Europe already had the first half of this lesson. In a June 2025 French Senate hearing on procurement and digital sovereignty, Microsoft's French legal lead was asked under oath whether he could guarantee that French citizens' data would never reach US authorities. His answer was no, he could not. Under the Cloud Act, a US-headquartered provider must comply with a lawful US order regardless of where its servers sit.
So both exposures are now on the record, from two different US companies, in two different ways.
The same jurisdiction, two ways to lose control
Data. A US provider can be compelled to hand over what runs through it. (Cloud Act, French Senate testimony, June 2025.)
Access. A US provider can be compelled to switch off what you depend on. (Export control directive, Fable 5 and Mythos 5, June 2026.)
The data center being in Frankfurt or Paris does not change who the provider must obey.
This is not an anti-American point
The US government may have acted in good faith. Security directives are not inherently unreasonable, and Anthropic itself supports a government's right to block genuinely unsafe deployments. For a European buyer, the problem is not motive. It is dependence. When your stack runs through one jurisdiction, you inherit that jurisdiction's politics, its security reviews, and its Friday-evening letters, whether or not any of it was aimed at you.
What changes when the model is open and the hosting is European
Here is the honest part, because it matters more than any pitch. Melious has no Fable, no Mythos, no GPT-5, no Claude. We are not the frontier. If your use case needs exactly the frontier capability that just went dark, we are not your provider this quarter.
What we run is different by design: 60+ open-weight models (Llama, Qwen, Mistral, DeepSeek) behind one OpenAI- and Anthropic-compatible API, hosted on European infrastructure with data under EU law. Open weights have a property that matters here. Once they are downloaded and running on EU servers, there is no central switch a foreign government can flip. You cannot recall a model that is already in the open and already serving traffic outside your reach.
We will not oversell it. This is not full immunity from geopolitics, and export rules can still touch model weights at the source. But the specific failure that hit Fable 5 users, a remote off-switch held by a single government, is one that open weights on European infrastructure do not have.
Run on infrastructure no single letter can switch off
60+ open-weight models, OpenAI- and Anthropic-compatible, hosted in Europe under EU law. No training on your data.
The take
The Fable 5 users who lost access on Friday did nothing wrong. Neither, arguably, did Anthropic. The model was simply on the wrong side of a border when a directive arrived, and that had nothing to do with how good it was.
If a government you did not elect can switch off your AI overnight, it was never really your AI.