On Friday, one of the most capable AI models on earth vanished in a matter of hours. Not deprecated. Not rate-limited. Switched off, worldwide, by government order. If your business depends on a single AI model, that is the most important thing that happened in technology last week, and the lesson it teaches outlasts the news by years.
Here is what happened, briefly and accurately. Anthropic released Fable 5 and Mythos 5, its two most capable models. Three days later, on 12 June 2026, the US government issued an export control directive barring access by any foreign national, inside or outside the country, including Anthropic's own foreign-national staff. Anthropic could not enforce that selectively, so it disabled both models for every customer, everywhere. Hundreds of millions of users lost access overnight. The company is contesting the order, says the capability the government is worried about is widely available in other models, and is working to restore access. It may well succeed. Its other models kept running.
So this specific outage may be temporary. The precedent is not.
The lesson that outlasts the news
Strip away the politics, the jailbreak claim, and the question of who is right. What remains is simpler, and more uncomfortable for anyone who builds on this technology.
A frontier capability that millions of people relied on became unavailable in hours, for reasons outside the customer's control and outside the vendor's control. Anthropic did not choose this. Its customers could not have seen it coming. The off switch was pulled by a third party that answers to neither.
Single-vendor, single-model dependence has been the quiet default for two years. One API key to the best model available, wired straight into the product. It felt like progress, and in capability terms it was. Friday was the first time most teams saw, in real time, what that dependence costs when the model stops working for reasons that have nothing to do with them. This is not the same as a normal outage, where you wait and your provider fixes it. Here, the provider was as powerless as you were.
Where this goes from here
Expect the conversation to widen quickly, and not return to where it was.
AI sovereignty stops being a conference panel and becomes a procurement question. Expect renewed energy behind European initiatives, sovereign model efforts, and data residency rules, driven less by ideology than by the simple wish not to have a critical input controlled from another jurisdiction. Open-weight and locally run models stop being framed as the cheaper, weaker option and start being valued for the one thing they offer that a hosted frontier model cannot: nobody can switch them off from a distance. Chinese labs, already strong and largely outside US export reach, become more attractive to buyers who want a hedge.
And the uncomfortable long arc is a bifurcation of the model landscape. National-only access for the most capable systems. Industry-gated access, where cyber, pharma, biotech, or defence get capabilities the general market does not. The era of everyone, everywhere, sharing one best model may turn out to have been a phase, not the destination. If you have read why AI is about to split the job market in two, this is the same fracture line, one level up the stack.
What to actually do about it
None of this means panic, and none of it means abandoning the frontier labs. That is still where the best capability lives, and for most businesses they remain the right default. It means you stop building as if availability is guaranteed, and you treat model dependence with the same resilience discipline you already apply to your database and your payment provider.
Three concrete moves, in order of effort.
First, put an abstraction layer between your application and the model. If switching providers means a rewrite, you do not have a strategy, you have a hostage situation. This is the same principle behind the advantage being everything you wrap around the model, applied to resilience rather than differentiation.
Second, keep a tested fallback from a different provider. Not listed in a document, actually wired in and exercised, so that failover is a config change and not a fire drill. The boring resilience work that looks like waste right up until the morning it isn't.
Third, hold an open-weight model you can run on infrastructure you control, as a capability floor that no outage and no directive can take from you. It will not match the frontier on the hardest tasks. It does not need to. It needs to keep your product alive while you wait for the headline to resolve. For a fuller treatment of when to own versus rent capability, see you don't need to build a brewery to drink a pint of beer.
Do not overcorrect
There is a failure mode on the other side. Optionality is insurance, not a religion. Spreading every workload across three providers to feel safe will slow you down, multiply your costs, and dilute the very advantage you are trying to protect. The point is not to distrust your main provider. It is to make losing them survivable.
The right posture is the one any good engineering leader already knows from every other critical dependency: use the best tool for the job, understand exactly what breaks if it disappears, and make sure that what breaks is an inconvenience you have planned for rather than an outage that ends your week. If you want help thinking through where your real concentration risk sits, and what proportionate resilience looks like for a business your size, that is exactly the kind of decision a fractional CTO is for.
Friday was a fire drill that happened to be real. The teams that treat it as one, and quietly build the optionality they were missing, will barely notice the next time. The ones that file it under interesting news will find out the hard way that the most important property of your best AI model is not how clever it is. It is whether it is still there tomorrow morning.
Be honest with yourself: if your best model disappeared before breakfast, how long until it broke something in production, and how long until you recovered?
If you are mapping your AI dependencies and want a second pair of eyes on where the real single points of failure are, I am always happy to compare notes. Let's talk.
Related: If your AI strategy is just ChatGPT, you don't have one · You don't need to build a brewery to drink a pint of beer · Agent sprawl is the new shadow IT. You need a control plane · AI is about to split the job market in two
