Refresh copy for 4M SKUs

Context

A retailer regenerates product copy for ~4M SKUs each season. Every call carries a large, identical brand-voice + policy prompt; caching that shared prefix collapses the input-token bill.

The decision

Caching is the build-vs-buy lever — with a high shared-prefix cache share the per-call cost drops enough to make the feature economically obvious.

What most miss

Teams price this at the naive per-call rate and conclude it's too expensive. The brand prompt is most of the tokens — cache it and the economics flip.

Stakes

Priced without caching, a 4M-SKU refresh looks unaffordable and the feature dies in the business case.