Examines how teams use Claude Code and finds that most of the cost surprises aren’t coming from where people expect. Rather than conversation history being the primary driver, other factors like context bloat merit attention, and the instinct to keep prompts tight or switch to smaller models may miss the actual cost sources.