Your AI access is not property. It is a permission that can be revoked.

June 12, 2026, 5:21pm ET. The US government sends Anthropic an export control directive. The effect: shut down Fable 5 and Mythos 5 immediately. For every customer in the world.

The key phrase was not "customers". It was "foreign national Anthropic employees" - Anthropic staff who are not US citizens. Even if they sit in an office in San Francisco.

Read that again. The state just said access to the best model can be cut by nationality. That is not software logic. That is ITAR logic - the logic of arms export control.

What actually happened

Anthropic says it is a misunderstanding. That this is a narrow jailbreak that "other publicly available models can find too". They are probably right about the technique.

But this was never about the technique. It is about the precedent. Frontier AI stopped being treated as a product. It became controlled strategic capability.

One commentator put it bluntly: API access is not your property. It is "revocable permission inside a sovereign-controlled stack".

This is not the first crack

Anthropic spent years building legitimacy. They were the "trustworthy" lab. The underdog with the cute mascot, publishing its safety work.

Then, in a matter of weeks, a series of moves broke that narrative:

• They blocked third-party harnesses - Claude Code stopped working in OpenCode, Hermes and others. The message to the ecosystem: this is not a place you can freely build.
• Silent degradation of Fable - the model was quietly weakened for AI researchers, biologists, STEM people. A doctor could not even say "hello" to the model, because the system knew he was a biomedical researcher. Anthropic walked it back after the storm on X. But as Dean Ball wrote: it is one thing to refuse transparently, quite another to train an AI smarter than you to deceive you.
• Expanded data retention - 30 days of prompt storage, without a paired promise not to train on it. Naveen Rao called it a red line: prompts contain our IP, our design files, our docs.

Legitimacy works like Fable tokens. You earn it over years and burn it in a week.

There is one lesson for Europe

Let's not kid ourselves. If you build a company in the EU on someone else's frontier API, you do not own your AI-engineering. You own a permission. And a permission can be switched off by someone in Washington at 5:21pm on a Friday.

The point is not that Anthropic is "evil". The point is that a sovereign state always beats a license to use. Always. Your enterprise contract is weaker than an export directive.

And here is the core: independence will be the main axis of software purchasing in the EU. Not price. Not feature count. The certainty that it still works tomorrow.

Why a smaller model you own beats a bigger one you rent

Reverse-engineer this. What do you actually want? A company that runs - regardless of what the US government or any lab does.

That means you need AI-engineering nobody can cut off. That means models you control. That means a stack you can run on your own hardware if you have to.

The choice is not "smartest model on earth versus weaker model". It is a model that is yours versus a model that is someone else's. Mythos codes everything - until someone turns it off.

Better a slightly weaker model tuned to you and your AI-engineering than a generic "Mythos does it all" that is revocable permission.

What Open Mercato is betting on

I'll say it plainly, because it sounds a bit political, but these are the facts.

Open Mercato is betting on independent AI-engineering for companies in Europe. Concretely:

1. People and code from the EU. A strong commitment to hiring developers and accepting contributions mainly from the European Union.
2. Open source models. We believe in Open Models you can run and own.
3. A Harness AI built for smaller models. For a while now we have optimized our Harness AI to build enterprise apps without losing quality - including on smaller models and models running on-device.
4. Removing legal and security barriers. So a corporation with compliance constraints can use AI without the risk that someone cuts off its access.

They may cut your Mythos. But that does not stop your programming with us. Your AI-engineering stays yours.

What you can do tomorrow

Take your AI stack and ask one question: what happens to my company if this exact model disappears on Friday at 5:21pm?

If the answer is "I'm stuck" - you don't have AI-engineering. You have a permission.

And a permission is not an asset. It is work on someone else's lot.