Three habits for the weekend when Fable 5 stopped working

I was prompted to change model while firing tasks to Fable 5 over the weekend, then I had to learn what had happened.

On 12 June 2026, not at the end of a deprecation notice, since the model had only just launched, and not because of an outage, the US government issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees. To comply, Anthropic disabled both models for every customer.

Luckily, I was only testing Fable 5 recently, with no deadline to chase, but if your product ran on the two models that afternoon, your Plan B might not be that obvious: you must have switched to Fable or Mythos for a reason, and that reason now needs answering somewhere else.

What is hard to argue with

I am not here to argue who is right.

The government cited national security, which is a serious basis. Anthropic has said it is complying while disagreeing, believes the matter is a misunderstanding, and is working to restore access, which is reasonable. Those details will be argued over by people far closer to them than most of us are.

The operational fact underneath, though, is hard to argue with: a widely used, commercial model, relied on by businesses around the world, can become unavailable overnight for reasons outside any contract, any amount of spend, or the vendor’s product roadmap. The directive was certainly outside anyone’s calendar too.

Lock-in is real, and not unique to AI either. So the lesson repeats. Maybe it is time to pick up some habits.

Three habits worth picking up

Treat vendor and model evaluations as a routine, not a one off event. Keep a short list of two or three credible providers and usable models, even lower capability ones, test them against your own tasks on a regular cadence, and design your system so that changing the underlying model is a contained change rather than a rebuild. An abstraction layer will not make a swap free, since prompts and quirks differ between models, but it turns a sudden loss into a manageable task rather than a crisis.
Own the parts that are truly yours. This directive is a sharp reminder that the question is not only commercial. Its stated scope was foreign nationals, inside and outside the United States, and the practical result was that the models went dark for everyone. Therefore, access to a frontier model can depend on a decision made in a jurisdiction that is not your own. You do not need to train a foundation model to reduce that exposure. The durable assets are your data, your evaluation sets, your prompts and scaffolding, and your understanding of your users or business problems. For most teams that is all you need to do. Where it genuinely matters, it can extend to running open models you control. Some call this AI sovereignty; the plain version is keeping enough capability in house that your operations are not wholly dependent on a single external provider.
Have a Plan B, and put it in production. For each critical use of AI, know which alternative you would move to, roughly how long the move would take, and what it would cost, then try that path occasionally rather than assuming it works. Better still, run it in production, side by side with your main pipeline. The Fable 5 and Mythos 5 episode is the argument itself: the notice may come close to a Friday afternoon at 5.29PM, and the teams who kept running were the ones who already had somewhere else to go.

None of this is a reason to avoid good tools, or to bet that access will not return. If anything, it is a reason to also lean on open source models you can run yourself. Anthropic says it is working to restore it, and it may well succeed.

By all means, use the best model for your users today. The point is to use it with your eyes open, knowing that any model can be withdrawn, whether by a roadmap, a price change, or a government directive.

Resilience is not built when the notice arrives. It is built in the ordinary months beforehand, by the teams who assumed this day would come.

Reader correspondence is welcome at me@xiongzhitao.me.