How we cut our LLM bill about 30% without making the product worse

Three levers that actually moved our token spend: caching what repeats, routing to the right size of model, and sending less in the first place.