Cut inference costs 60 percent by switching models and tracked margin impact

We migrated from GPT-4 to a distilled 7B model for our summarization endpoint last month. Cost-per-output dropped from $0.04 to $0.016 while maintaining 92 percent quality scores. Gross margin expanded from 55 percent to 78 percent without changing customer pricing.