How to Diagnose a Sudden LLM Cost Spike Using Six SQL Queries

When an AI language model bill unexpectedly doubles overnight, the cause typically falls into one of about eight categories that can be systematically identified through request log analysis. Before investigating, developers are advised to immediately cap spending via provider limits or rate controls to stop costs from rising during diagnosis. A structured runbook recommends running six sequential SQL queries, each designed to eliminate a possible cause, starting with whether the spike stems from higher request volume or increased cost per request. Key database columns such as attempt and model—captured from the response rather than the request—help detect retry storms and silent model alias changes. The guide also stresses avoiding simultaneous fixes, as disabling multiple suspects at once may resolve the incident but leaves the root cause unidentified and likely to recur.