Claude Code's 33K-token startup tax dwarfs OpenCode's 7K

Before an AI coding agent reads a single word of your actual request, it front-loads a fixed payload — system prompt, tool schemas, and environment context. The benchmark behind this comparison puts Claude Code’s baseline at roughly 33,000 tokens against about 7,000 for the open-source OpenCode, a nearly five-fold gap in fixed overhead per invocation.

That overhead isn’t paid once; it rides along on every turn. A heavier baseline directly inflates per-request cost and time-to-first-token, consumes context-window headroom that could otherwise hold code, and can dilute the model’s attention with boilerplate before the real task even begins. The usual drivers are an expansive set of built-in tool definitions and a verbose system prompt.

The tradeoff is the interesting part: the larger prompt likely buys capability — more tools wired in and richer default instructions — while OpenCode’s leaner footprint favors efficiency. The comparison surfaces a real design tension between out-of-the-box power and token thrift. Worth treating the exact numbers as a single-source snapshot rather than a fixed spec, since prompt sizes shift across versions and configurations.