I benchmarked 8 LLMs for a niche production app. The flagship cost 5.8x more - and lost.

My app generates personalized readings for BaZi — Chinese "Four Pillars" birth charts. Every reading is an LLM call, every call costs money, and the domain is full of trap terminology that models love...