A user analyzed 185 transcripts from Claude Code sessions to understand token consumption, discovering that tool output accounts for 98.7% of context usage, while user prompts make up only 0.7%. The analysis revealed that each sub-agent spawn incurs a significant token cost, roughly 68,000 tokens, which only becomes cost-effective if the delegated task is substantial enough to offset this re-purchase of context. The user has shared their Python scripts on GitHub to enable others to analyze their own Claude Code logs. AI
IMPACT Provides insight into LLM operational costs and efficiency for users, highlighting the impact of tool output on token consumption.
RANK_REASON User analysis of a specific product's token usage, not a product release or official benchmark.
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
