Classify the task
Extraction, classification, reasoning, rewriting and critique do not necessarily need the same model.
Route by risk
High-stakes synthesis may justify a stronger model, while deterministic formatting or simple classification may not.
Add fallbacks deliberately
Fallback models should preserve output contracts and be tested rather than activated as an emergency afterthought.
Measure end-to-end economics
Cost per successful workflow matters more than cost per individual model call.