The cheap/expensive pairing- A way to reduce token usage in models :-P

A recurring, valuable pattern: offer a cheap, narrow variant alongside an expensive, broad one, so the model can pick the right cost for the task.

• "Find this phrase in the document" (cheap, targeted) vs. "read the whole document" (expensive, complete).
• "Get just the metadata" (cheap) vs. "fetch the full text" (expensive).
• "List what's available" (cheap discovery) vs. "fetch these N items" (expensive retrieval).

Without the cheap option, the model pays full price for every lookup: more tokens, more latency, more cost. The cheap variant lets it confirm a fact or locate a needle without ingesting a haystack. Give the model economical options and it will often use them.