Claude Code Sends 4.7x More Tokens Than OpenCode Before Reading Your Prompt
- source_url: https://systima.ai/blog/claude-code-vs-opencode-token-overhead
- source_type: article
- platform: blog
- author: Systima
- original_date: 2026-07-12
- added_date: 2026-07-22
- category: coding
- tags: harness, token-overhead, prompt-cache, claude-code, opencode, cost
- quality_score: 5
- hn: https://news.ycombinator.com/item?id=48883275
摘要（中文）
在 API 边界挂 logging proxy，同模型同机同任务对比 Claude Code 与 OpenCode。空仓回 OK：CC 首轮约 33k token 地板 vs OC 约 7k（工具 schema 为主）。OC 前缀跨 run byte-identical；CC 中段 cache rewrite 可使 cache-write 最高约 54×。72KB 指令文件约 +20k/请求；小 MCP 每服约 1–1.4k；双子 agent 同任务可从 121k→513k（约 4.2×）。多步任务上 batching 有时抹平地板，但换模型后优势可消失。读法：先量地板/轨迹/缓存稳定性，再比单价。
Summary (English)
API-boundary logging proxy study of Claude Code vs OpenCode on the same model, machine, and tasks. Empty-workspace 'OK' floor: ~33k tokens for Claude Code vs ~7k for OpenCode, dominated by tool schemas. OpenCode prefixes were byte-identical across runs; Claude Code rewrote cache mid-session up to ~54× more cache-write tokens on the same task. A 72KB instruction file adds ~20k tokens per request; modest MCP servers add more; two subagents multiplied a 121k task to 513k (~4.2×). Multi-step batching can erase the floor advantage depending on model behavior. Measure floor, trajectory, and cache stability before comparing $/M.
One-liner
同模型换壳账单差在 harness 地板、请求轨迹与 cache 前缀是否稳定。
Source body / metadata
Fetched via opencli web read / official docs during evening intake. Key claims are grounded in the source page and same-day Obsidian digest notes.
API-boundary logging proxy study of Claude Code vs OpenCode on the same model, machine, and tasks. Empty-workspace 'OK' floor: ~33k tokens for Claude Code vs ~7k for OpenCode, dominated by tool schemas. OpenCode prefixes were byte-identical across runs; Claude Code rewrote cache mid-session up to ~54× more cache-write tokens on the same task. A 72KB instruction file adds ~20k tokens per request; modest MCP servers add more; two subagents multiplied a 121k task to 513k (~4.2×). Multi-step batching can erase the floor advantage depending on model behavior. Measure floor, trajectory, and cache stability before comparing $/M.
Obsidian evidence
- local_note: 调研/2026-07-22-Coding-Agent-Harness预载税与缓存失稳.md
- intake_run: daily-intake-evening 2026-07-22