Mythos and Fable's regional cutoff put model availability on Europe's product risk register

The user-visible version of a model cutoff is dull. A button that used to write the first draft returns an error, and the ticket that lands says the AI is broken again. Nobody outside the building learns which vendor drew which line across which map.

Inside the building, the cost sits in everything that was tuned. Prompts get shaped to one model's refusal habits and its tool-calling format. The eval set that told you the feature was good enough was scored against that model's outputs, so it certifies a model rather than a capability. Re-pointing a client library takes an afternoon. Re-earning the confidence you had in the output takes as long as your eval cycle takes, and most teams have never timed that cycle when it was not an emergency.

The version I keep hearing is: we sit behind a router, we support four providers, we can switch. The version I keep finding is one provider's prompts in production, a second provider's credentials sitting unused in a secrets manager, and a fallback path last exercised in staging by an engineer who has since moved on.

Two details in TechCrunch's account stop this being a tidy European grievance. The line was drawn around Europe rather than through it ^[1], so the teams treating it as a warning were reading someone else's outage, and reading it correctly. And the account is thin on specifics: it gives no stated reason for the withdrawal, puts no date on it finer than earlier this year, and does not say whether access outside Europe came back ^[11]. What TechCrunch reports as the measured blast radius across the ecosystem is two executives quoted, with no count given for how many teams were affected ^[12].

The concentration arithmetic is worth doing once. Two named models went away in what the source describes as a single change of availability, which means a team that had thoughtfully split its workloads across Mythos and Fable was carrying one point of failure, not two ^[10]. Diversity of model names inside one lab buys nothing here.

The grid I would draw has swap cost on one axis, measured in hours to re-point plus hours to re-run evals, and blast radius on the other: does one feature degrade, or does the product stop working. Only the high-cost, high-radius quadrant justifies a warm second path, meaning prompts maintained for both models and a scheduled hour where the primary provider is switched off in production on a working day. Low radius gets a written plan and nothing else, because buying live redundancy for a nice-to-have summary is how a platform budget gets spent before the thing that matters breaks.

During that hour, the numbers to watch are task completion and time-to-value on the affected flow. Session counts will barely move, because people keep opening the app whether or not the feature answers, which is exactly why engagement figures will tell you nothing about what you just lost.

This is for the person who signed the data processing agreement and assumed it covered this. That document governs how data is handled, not whether a model answers a request from your region. Meredith Whittaker's line at the conference about sovereignty concerns producing paying customers ^[6] and Emad Mostaque's definition of sovereignty as the ability to resist power being exerted over you ^[7] land, for a product team, as the same small operational question: how many hours between losing a model and trusting its replacement, and when did you last measure it.