When Fable 5 went dark, the companies that felt it worst weren't the ones using Claude to write emails. They were the ones whose product was, underneath, a thin layer over Anthropic's model. One letter, ninety minutes, and their core feature started returning errors, with nothing they could ship to fix it. Their roadmap was suddenly a negotiation between Washington and a company they don't work for.
In Part 1 I argued the model in your stack isn't yours. If that's true, the question is what you should have built instead.
Start with the failure that just got demonstrated. Snyk, looking at the suspension through a security lens, said it plainly: a single directive took a generally available product offline for its entire global user base within hours, which makes model availability "revocable by forces beyond their and their vendor's control." Their name for what every team should now map is the "AI blast radius" - everything in your product that breaks if one model disappears. Mark Carney reached the same conclusion from the other end of the telescope, warning Canada about "overreliance on a limited number of American AI providers." A security engineer and a prime minister, same week, same lesson: the model is a dependency you don't control, and most teams priced it like one they did.
Most teams have this backwards. The raw model was never the moat. Everyone can rent the same one, on the same terms, at the same price, your competitor included. Whatever advantage you actually have lives in the layer you wrap around it: your context, your evaluations, the workflows you've encoded, the institutional knowledge you've turned into something a machine can operate. That layer is yours. The model underneath is a commodity with a kill switch.
Build as if the model can vanish, because it just did. Map your AI blast radius. Keep more than one model qualified - closed and open - so a swap is a config change rather than a six-month migration. Your own evaluation set is the other half: it's what lets you certify a replacement in an afternoon instead of trusting a vendor's benchmark. Context, memory, and traces should live in your systems, not locked inside one provider's account, and you route by the capability you need rather than the brand on the API key.
I'm not claiming the lockdown wins in the end. Open-weight models from DeepSeek, Kimi, and others, plus the price war running underneath all of it, push hard the other way, and these controls may erode as fast as they appeared. That uncertainty is the whole reason to hold optionality. You don't want your company wagered on the closed frontier staying available, or the open one staying good enough. You want to be able to move.
This is the bet we made building Plexi Flexor. The model is the least interesting part of the system. The skills, the memory, and the context that make an AI useful for your business sit in a layer above any single model, and it routes by capability tier rather than by vendor, so the thing underneath is a setting you change, not a foundation you pour. The afternoon Fable went dark is the reason that layer exists. It's the difference between editing a config and apologizing to your customers.
The model is the easiest piece of your stack to replace. Everything you built around it is the hard part, and the valuable one. If your most important model vanished tonight, how much of your company would still be standing in the morning?
This is Part 2 of 2. Start with Part 1: A government switched off a frontier AI model in an afternoon.
