Claude Code Uses 33,000 Tokens Before Any Query… Real-World Usage Reaches 75,000 Tokens

Anthropic's AI coding tool 'Claude Code' transmits tens of thousands of tokens to the AI model before processing user instructions, according to an analysis showing real-world usage can reach up to 75,000 tokens. A comparison experiment by AI consulting firm Systima revealed significant differences in token consumption between Claude Code and open-source tool OpenCode. In a minimal environment, Claude Code sent 32,800 tokens for a simple 'print OK' command, while OpenCode used just 6,900 tokens. The most pronounced disparity occurred in tool description volumes, with Claude Code transmitting 24,000 tokens for 27 tool descriptions versus OpenCode's 6,900 tokens for 10 descriptions. Even with all tools disabled, Claude Code's system prompt required 6,500 tokens compared to OpenCode's 2,000. This pre-emptive token usage consumes approximately 1/6 of Claude Sonnet 4.5's 200,000 token context window before any code or user input is processed. While prompt caching reduces retransmission costs, it does not reduce the number of context tokens. Project rule files like 'CLAUDE.md' also add significant overhead, increasing token usage by 20,000 per request.