Cost at peak, not average
Budgets estimated on average call volume break during campaigns or batch jobs. We size for peak traffic and retry rates, and keep a cheaper fallback model in reserve.
Control means output boundaries
Is the same input stable, does the model refuse malformed inputs, can sensitive fields leak? Run real samples before launch instead of trusting a demo.
Replaceability is insurance
Models get updated, prices change, endpoints retire. Wrapping calls in a thin adapter and keeping prompt versions plus regression samples is what makes a swap a one-day job.
One lesson
We would rather take the second-best model that we can swap out than the best model we are locked into. Businesses run for years; models change more than once.