Well, I heard of one case where a bug was causing queries of 100,000 tokens at a time. https://medium.com/dataherald/cutting-llm-costs-by-83-with-l...