681 AI Sessions Tracked: Repeated File Edits, Not Heavy Use, Drain Your Budget

A developer tracked 681 AI work sessions across 41 projects over four months, from April to August 2026, to understand where subscription costs actually go. The analysis found that 90% of requests cost very little, while just 15% of requests — those where the AI edits the same file three or more times — account for 41% of total output. When the AI loops on a problem without progress, it generates six times more text than a normal request while rarely improving on its second attempt. A key behavioral signal is the phrase 'let me try something else,' which the author identifies as the point where the AI stops reasoning and begins cycling through repeated solutions. The author recommends setting a two-attempt limit per problem in the AI's instruction file and notes that re-explaining project context costs time but not significant money.