What happened
Systima.ai measured Claude Code's token consumption against OpenCode, finding Claude Code uses substantially more tokens for system prompts, tool schemas, and injected scaffolding. For a one-line reply, Claude Code consumed approximately 33,000 tokens before the user's prompt, compared to OpenCode's 7,000. Claude Code also wrote up to 54 times more cache tokens mid-session for the same task, indicating cache inefficiency. Configuration files and subagents further inflate Claude Code's usage, adding around 20,000 tokens for a 72KB instruction file and quadrupling costs for tasks fanned out to two subagents. Claude Code can achieve lower total token counts on multi-step tasks by batching tool calls.
Why it matters
Increased token consumption directly translates to higher operational costs and reduced context budget for agentic AI deployments. Procurement teams face elevated billing; platform engineers and architects manage constrained context windows. Claude Code's 54x higher cache writes and 20,000-token overhead for a 72KB config file demonstrate significant cost multipliers. This follows prior reports of LLM coding costs exceeding subscriptions tenfold, underscoring the need for granular token usage monitoring. Teams should prioritise tools with efficient token management to preserve context and control expenditure.
