On 12 June, two of the most capable AI models on the market went dark — not because the company that built them chose to retire them, and not because a customer cancelled a contract. They went dark because a government ordered it.
Anthropic disabled Claude Fable 5 and Mythos 5 after the US government issued an export-control directive citing national security. The order suspended access for any foreign national, inside or outside the United States. Because no provider can verify nationality in real time at the point of use, the only way to comply was to switch the models off for everyone. Anthropic's other models kept running. These two did not.
Read the company's own account and the picture gets sharper. The directive arrived late on a Friday afternoon and named no specific concern. The trigger, as Anthropic understands it, was a demonstrated technique for getting the model to read a codebase and flag software flaws — something the company says other public models already do without any bypass at all. Anthropic is complying. It also disagrees, calls the action a misunderstanding, and is working to restore access.
You can argue the politics of that all week. Most of LinkedIn already is. I want to talk about something less dramatic and more useful — what it means for anyone who has wired a hosted model into how their business actually runs.
The category error
Here is the uncomfortable part. Most organisations treat AI model access as procurement. You pick a vendor, you sign for a tier, you integrate the API, you move on. Procurement assumes that what you bought is yours to use under the terms you agreed.
This event breaks that assumption. The model did not disappear because of anything the customer did or failed to do. It disappeared because a third party — neither the buyer nor the seller — reached in and pulled it. No service-level agreement covers that. No contractual remedy exists for a government export order. The thing you depended on answered to an authority that was never party to your deal.
That is not procurement. That is supply chain. And supply-chain risk has a discipline of its own — one that most AI adoption has skipped entirely.
The realistic response — and the fantasy one
So what is the sensible reaction? Not the one the loudest voices are selling.
The fashionable answer is sovereignty — build your own model, own the whole stack, escape dependency entirely. For a handful of states and a few very large enterprises, that conversation is real. For everyone else it is a fantasy. You are not going to train a frontier model. You should not try.
The defensible version of self-reliance is narrower and entirely achievable. It is not about owning a model. It is about owning your implementation — and refusing to let any single model become load-bearing.
Three things make this possible now in a way it was not eighteen months ago.
The first is abundance. Capable models are arriving from every direction — American labs, European labs, Chinese labs, open-weight releases you can run on your own hardware. The monoculture is breaking. When one source reaches a given capability, several others are not far behind, and for most real-world tasks the gap between the frontier and the merely excellent keeps shrinking.
The second is abstraction. Routing layers — OpenRouter is the obvious example — sit between your application and the models, exposing many providers through a single interface. You write your integration once. You change the model behind it with a configuration line, not a rebuild. Lock-in stops being structural and becomes a choice.
The third is portability by design. If you test your core business case against several models from the start — not one, several — you learn what genuinely depends on a specific provider and what does not. You find the prompts that need rework. You build the muscle of switching before you are ever forced to switch in a crisis.
Do those three things and a single off-switch, wherever it sits, loses its power over you. It becomes an inconvenience, not an outage.
The honest cost
I want to be direct about the price, because the people selling sovereignty rarely are.
Portability is not free. Models are not interchangeable parts — the same prompt produces different results across them, and "swap the model" rarely means "identical output." Testing across providers is real work. Running open-weight models yourself trades frontier capability and managed convenience for control and continuity, and it hands you a maintenance and security burden you were previously outsourcing to someone else.
So this is not a call to localise everything. It is a call to decide. Which workloads can your business genuinely not afford to lose for a week? Those are the ones that justify the cost of independence — multi-model fallback, an abstraction layer, local deployment where continuity matters more than raw capability. The rest can stay on whichever hosted frontier model is best today, precisely because losing it would be survivable.
That decision — survivable versus not — is the whole exercise. It is also the part almost nobody has done.
What actually changed this week
The Fable 5 suspension will be resolved, probably soon and probably quietly. The precedent will not be un-set. Whoever holds that switch — this government or another, this model or the next — has now demonstrated a willingness to use it.
The lesson is not to predict the next intervention. It is to build so that the next one does not matter. Treat model access as supply chain, not procurement. Own your implementation, not the model. And make sure the question "what happens if this disappears on Friday" already has an answer you have tested — not one you are improvising on Saturday morning.