🔗 Claude Code sends 4.7x more tokens than OpenCode before reading your prompt

Claude Code's appetite for tokens far exceeds that of OpenCode, as the article reveals startling figures from direct comparisons. In experiments run under identical conditions, Claude Code averaged around 33,000 tokens for a one-line reply, while OpenCode only consumed about 7,000. The data reported showcases how Claude Code’s approach introduces significant cache inefficiencies, requiring 54 times more tokens for some tasks. A primary concern raised is the impact on production environments, particularly when operating under regulations like the EU AI Act. The emphasis is on understanding the tokens spent is crucial for optimizing agentic AI systems, where every token represents context that can't be utilized for task completion. Additionally, analyzing not just input but also cache writes reveals the complexity and potential costs associated with these AI harnesses.