What Happens When the Model You Built On Disappears

What happens when the model you built on disappears.

Fable and Mythos just got shut down by government action. Not deprecated, not sunset with a migration path. Shut down. Anyone who wired those into a production workflow woke up to broken systems and no recourse. An unfortunate reality of using rented AI models in your workflow. When the model lives behind someone else’s API, you are renting capability you do not control, and the lease can be cancelled by a regulator, a policy change, or a quarterly business decision.

I run a hybrid stack precisely because of this. Most work does not need a frontier model. Drafting, summarization, classification, code completion, the long tail of routine tasks all run fine on local models I own outright. On Apple Silicon and comparable hardware the unified memory makes this practical today, not in some future roadmap. The data never leaves the machine, the marginal cost per token is effectively zero, and the latency is better because there is no network round trip. For the narrow slice of work that genuinely needs the frontier, I route out to the API and pay for it deliberately.

That is the architecture I think most serious teams land on. Not local instead of frontier. Local for the eighty percent, frontier for the twenty that requires it.

The framing that matters is sovereignty. Owned weights cannot be shut down. They cannot be repriced overnight. They do not vanish because a vendor changed terms or a government changed its mind. You can audit them, pin a version, and run them in five years exactly as they run today.

Open weight models keep closing the gap with the frontier tier, which means the slice that requires a rented model keeps shrinking. Build so that when the next API goes dark, and one will, the lights stay on.