A user discovered that 98.1% of the tokens processed by Claude Code over a 30-day period were for cache reads, with only 0.37% attributed to the model's actual output. This finding emerged after the user built a tool to track token usage, revealing that each request resends the entire conversation history, which is then served cheaply from Anthropic's prompt cache. The user also detailed how Claude Code stores session data in JSONL files and provided a Python script to accurately count tokens, avoiding overcounting due to how replies are logged. AI
IMPACT Highlights potential inefficiencies in LLM token management and cost structures for users.
RANK_REASON User-generated analysis of a product's operational details.
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →