Changes that cut our LLM pipeline costs more than model-switching did (opens in new tab)

I have been building multiple LLM systems and for our Organization biggest cost savings weren't from prompt-wordsmithing or model switchings. Sharing useful to anyone watching their token bill :