Last week we wrote that your AI model is a leased component on the vendor’s depreciation calendar. This week the lease got shorter. On June 9, Anthropic released Claude Fable 5, a guarded, public version of its otherwise restricted Mythos-class model. Three days later, after a US export-control directive, Anthropic suspended access to both Fable 5 and Mythos 5 for all customers. Not on a deprecation schedule. Overnight, by an authority that was neither the vendor nor you.
Why It Matters
Last week’s risk was the vendor’s calendar, a model retired on 60 days notice. This is the same dependency with the notice removed. The order nominally targeted foreign nationals, but Anthropic pulled both models worldwide to comply, then publicly disagreed. The company that built the model could not keep it online, all over a narrow jailbreak Anthropic says exists in other shipping models too. If one security finding can recall a tool made available at massive scale, the off switch for your workflow may sit with people you will never meet.
The Decision
This was a frontier model under a national-security order, not your everyday CRM feature. But the mechanism is identical, a tool you depend on kept online under someone else’s authority. The frontier just showed the shape first.
Call it the Off-Switch Dependency Test. If a tool can be removed by someone outside your company, it is not just software. It is an external dependency. The more revenue, decisions, or customer delivery it touches, the more fallback planning it needs.
So the question this week is not which model is best. It is how much of your operation runs on tools a third party can switch off without warning or your consent. A reasonable CEO can accept that exposure for the capability, or build the fallback first. Both are defensible. Pretending the switch is yours is not.
What To Do This Week
- Name the one AI tool your team would most miss if it went dark tomorrow morning.
- Ask its owner what the fallback is, and how many hours it takes to reach it.
- For anything load-bearing, write down who could pull it, a vendor or a regulator, and what you would do in the first hour.
What Not To Do
Don’t read this as a reason to avoid frontier tools. The capability is real, and the laggards will feel it. Don’t assume a bigger vendor means a safer supply. Size did not keep this model online. And don’t mistake convenience for control. You chose these tools. Underneath it all sits a deeper question. Are we, the users, quietly losing the freedom to keep the tools we choose?
Signal Boost
TechCrunch has the clearest narrative of the order and its business context. For the source of record, Anthropic’s own statement gives the suspension and its disagreement with the directive. The precedent is the point: Anthropic warns the same standard applied industry-wide would essentially halt new model releases.