Claude Code sends 33k tokens before reading the prompt; OpenCode sends 7k (systima.ai)

Recent testing revealed significant differences in token efficiency between Claude Code and OpenCode when processing requests. Claude Code consumed approximately 33,000 tokens before even receiving a user prompt, while OpenCode managed to keep its token usage around 7,000. This marked inefficiency stems from Claude Code's architecture, which often necessitates rewriting extensive cache tokens mid-session, leading to increased costs compared to OpenCode's uniform request prefix. As a consequence, users may find operational costs climbing when deploying Claude Code, particularly in production environments governed by regulations such as the EU AI Act, where understanding system behavior is crucial. Despite its higher baseline consumption, Claude Code did perform better on multi-step tasks by batching requests, resulting in lower total token usage compared to OpenCode, which processes one tool call per turn. The findings emphasize the importance of understanding token overhead in context budgeting and operational costs, especially as observed in subagent configurations that can amplify token consumption significantly. Overall, these insights into how different AI harnesses manage prompts are vital for developers and companies seeking cost-effective implementations in the evolving landscape of AI and machine learning.