Sending every task to the strongest model is like printing every memo on the most expensive paper. It works, but it's unnecessary.
01. Classify the Work
Three rough tiers are enough: mechanical work (classification, tagging, reformatting), middle work (summarising, drafting) and hard work (multi-step reasoning, code, analysis). These three call for different model classes.
02. Where a Small Model Suffices
Routing an incoming request into a category, translating, extracting fields from free text. Small fast models come close to large ones here at a fraction of the cost.
03. Where a Large Model Is Needed
Reconciling contradictory information, interpreting a long contract, building a multi-step plan. Here a small model's mistakes are silent: the answer looks plausible but is wrong.
04. Routing
A simple router can send a request to the cheap model first and escalate when confidence is low. On high-volume flows this alone cuts cost noticeably.
05. The Context Window Trap
A large context window is not an invitation to throw everything in. Irrelevant text lowers accuracy and raises cost. The right approach is finding the relevant passage and sending that.
06. Don't Choose Without Measuring
Build a small evaluation set of 50 to 100 examples from your own data. When you consider switching models, that set gives you a measurement instead of a guess.