A recent analysis of Claude Code usage reveals that the vast majority of tokens consumed are not from the AI's direct output, but rather from the conversation history being re-sent with each turn. The author's detailed count across 173 sessions found that direct output tokens constituted only 0.2% of the total, while re-sent context made up 97.5%. This means that the cost of any element within a session is its size multiplied by the number of turns it remains in the conversation, contrary to the common advice of simply asking the AI to be more concise. AI
IMPACT Understanding token consumption is crucial for managing costs and optimizing AI interactions, especially for developers using tools like Claude Code.
RANK_REASON Analysis of AI product usage by a third-party author.
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →