I compiled LLM inference pricing across 7 providers — the caching numbers are surprising(spreadsheet included) [R] — PLINKFEED