Our AI stack runs on frontier models, and I'm comfortable with that — they're better, and pretending otherwise is wishful thinking. But the dependency is real: pricing changes, deprecations, rate limits, terms you don't control. The standard answer is an abstraction layer so you can "swap models anytime." Necessary, and mostly theater — an abstraction you've never exercised against a real alternative is a comment in the code, not a capability.
So we keep an open-weight lane actually running: Qwen, served through Ollama, wired into the same interfaces as everything else. Periodically we re-run our eval set against it and update a simple map — which of our workloads it could carry today (classification, internal tooling, well-scoped scaffolding) and which it can't (the frontier-grade reasoning our products depend on). That map is the point. "We could switch" is a feeling; a current list of what would survive the switch is a plan.
Nobody is migrating. The frontier models earn their place daily, and the gap is real. But insurance isn't a bet that the fire happens — it's the option to stay calm if it does. This one costs us a machine and an occasional afternoon of evals. Cheap, for sleeping well in a single-vendor market.