Claude Code Burns 33k Tokens Before First Prompt
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Claude Code sends approximately 33,000 tokens of system prompt, tool schemas, and scaffolding before receiving a user's prompt, compared to OpenCode's 7,000 tokens
- Claude Code rewrites tens of thousands of cache tokens per request during a session, while OpenCode uses a byte-identical prefix, making its cache usage far more efficient and cheaper
- Instruction files like AGENTS.md or CLAUDE.md add roughly 20,000 tokens to every request in both systems, but Claude Code only reads the file when named CLAUDE.md, unlike OpenCode which accepts either name
- MCP servers add 1,000 to 1,400 tokens per server per request, with five servers adding nearly 7,000 tokens, and Claude Code silently ignores project-scoped .mcp.json unless explicitly configured
- Subagents in Claude Code multiply costs: a 121,000-token task ballooned to 513,000 tokens when fanned out, due to duplicated bootstrap overhead per agent
- Claude Code batches multiple tool calls into fewer requests, which on a multi-step task led to lower total tokens (121k) than OpenCode (132k), despite its higher per-request baseline
- Logging proxy measurements at the API boundary revealed exact payloads and metered usage, showing that baseline overhead consumes a significant portion of context windows before user input begins
Why it matters: Teams using Claude Code face higher latency, cost, and context pressure from the start of every session, especially with subagents or rich tooling. While batching can offset this on long tasks, the overhead makes lightweight or high-frequency operations significantly more expensive and less scalable than with leaner agents like OpenCode.
