The day the model disappeared: why I architected Kiln so no vendor's decision could stop my build
A provider made a frontier model unavailable to me overnight, no warning. It cost me twenty minutes, not a rebuild. Here's the architectural bet that made the difference.
A few weeks ago, in the middle of building something, a model I was relying on just… wasn’t there anymore. Overnight, a provider made a specific frontier model unavailable to me. No warning, no migration path — just gone.
It cost me maybe twenty minutes of confusion and a rerouted task. In the grand scheme, nothing. But it crystallized something I’d been building around for six months without fully saying out loud: you do not get to control the thing you don’t own.
I’d already made the architectural bet before this happened — Kiln, the inference layer underneath Foundry, was built from day one to treat every model as swappable. Not because I predicted this exact moment, but because I’d watched enough of the industry to know that some version of this moment was inevitable. A vendor changes terms. A price goes up 10x. A capability gets fenced off behind a new tier. A model gets deprecated with six weeks’ notice. It always happens to someone, eventually.
The instinct most people have is to treat this as a rare, unlucky event. I think that’s backwards. The rare thing is a system where it can’t happen to you.
So here’s the actual lesson, and it’s smaller and more useful than “diversify your vendors” (true, but vague): build the seam before you need it. Kiln doesn’t route to one model — it routes to a pool, with a compliance layer that decides who’s even eligible before quality or cost ever gets weighed. That seam isn’t a feature I added after getting burned. It’s the reason getting “burned” cost me twenty minutes instead of a rebuild.
If you’re betting real work on a single model, a single vendor, a single API — you don’t need to rip that out today. You just need to know where the seam would go, so that when your version of this happens, you’re rerouting, not rebuilding.
That’s the whole post. Not a complaint about anyone. A note to future-me, and to you, if you’re building on top of someone else’s roadmap: the seam is cheap to build in advance and expensive to build in a panic.