On Tuesday, Anthropic put its most capable model into public hands. By Friday, it had pulled it, to comply with a US government export control directive.
Claude Fable 5 launched mid-week, described as state of the art on nearly every benchmark, and shipped with safeguards built in that routed anything touching cybersecurity, biology or chemistry to a weaker model. Days later, Anthropic suspended access to Fable and its less restricted sibling Mythos entirely.
For a lot of teams building on the newest release, that is a scramble. For us, it was a non-event. Nothing we run changed.
We build client workflows on models that are stable, proven and boring, and we keep the model swappable. That is the whole idea behind how we work. The workflow is the asset. Which model sits inside a given step is a detail you should be able to change in an afternoon, without the work falling over.
So our response to a week like this is the same as our response to most AI news: note it, check that nothing depends on it, carry on.
There is a wider point here for anyone planning an AI rollout. The capability at the frontier is real, and so is the volatility around it. A model can sit at the centre of your workflow on Monday and be unavailable on Friday, for reasons that have nothing to do with you: a vendor decision, a safety review, an export rule. If your operation only works while one specific model is available, you have built on sand.
Readiness is not having the latest model. It is being able to lose one and keep moving. The rest is just chasing.
Infrastructure and Governance are two of the five Dimensions the Diagnostic covers, alongside Mandate, Process and People. Both come down to the same quiet discipline: building so that the parts you do not control cannot take you down.