The 33k-token preamble is a bet. Here's how to check if it's paying off.

Short summary

A logging proxy comparison reveals Claude Code sends ~33k tokens of scaffolding per first turn vs OpenCode's ~7k, but the mass is mostly tool definitions (27 tools vs 10), not verbose prompts. Claude Code loads a full orchestration platform with subagent delegation and worktree management; OpenCode loads a lean editor. The 33k preamble is a wager that extra capabilities earn back their token cost — sometimes it saves money on multi-step tasks, sometimes it quadruples the bill via subagent fan-out. Prompt caching behavior, not raw preamble size, determines steady-state cost.

• •Claude Code's 33k token preamble is mostly tool schemas (27 tools vs OpenCode's 10), not verbose system prompts
• •The preamble is a wager: extra capabilities can save tokens on complex tasks but quadruple costs on simple fan-out
• •Prompt caching behavior, not preamble size, is the real determinant of steady-state token cost

Generated with AI, which can make mistakes.