Claude Code vs OpenCode: Token Overhead Measured at 33k vs 7k
Original: Claude Code sends 33k tokens before reading the prompt; OpenCode sends 7k
Why This Matters
Token overhead benchmarks directly inform cost planning for teams running agentic AI coding workflows at scale.
Systima measured token usage for Claude Code and OpenCode on identical models and tasks. Claude Code sends ~33,000 tokens of system prompt and scaffolding before reading the user prompt; OpenCode sends ~7,000 — a 4.7x difference on Sonnet 4.5, narrowing to 3.3x on Claude Fable 5.
Systima conducted a controlled comparison of Claude Code and OpenCode, placing both harnesses on the same model, machine, and tasks, then inspecting all API-level traffic. On a minimal one-line task with Sonnet 4.5, Claude Code consumed roughly 33,000 tokens of system prompt, tool schemas, and injected scaffolding before the user prompt was even processed; OpenCode used approximately 7,000. Re-running on Claude Fable 5 narrowed the ratio to about 3.3x, as Anthropic ships a smaller system prompt to newer models.
Cache efficiency was a major differentiator: OpenCode's request prefix was byte-identical across every captured run, paying a single cache-write cost per session. Claude Code repeatedly rewrote tens of thousands of prompt-cache tokens mid-session, generating up to 54x more cache-write tokens than OpenCode on the same task — relevant because cache writes are billed at a premium.
Real-world configuration compounds the overhead further. A 72KB AGENTS.md or CLAUDE.md instruction file adds roughly 20,000 tokens per request. Five modest MCP servers add 5,000–7,000 more. A production setup therefore reaches 75,000–85,000 tokens before any user input arrives. Subagent fan-out also multiplies cost: a task costing 121,000 tokens directly ballooned to 513,000 tokens when split across two subagents, since each subagent carries its own bootstrap overhead and the parent consumes the full transcript.
One result favored Claude Code: on multi-step tasks, its ability to batch tool calls into fewer requests produced a lower whole-task token total than OpenCode, which re-pays its smaller per-turn baseline repeatedly. The study concludes the baseline starts higher for Claude Code, but session structure ultimately determines overall spend.