Summary：**Surprising: Claude Code Uses 33k Tokens Before Prompt, OpenCode Only 7k** *We measured what Claud

referrerpolicy="no-referrer"
style="max-width:100%;height:auto;display:block;margin:0 auto;">

**Surprising: Claude Code Uses 33k Tokens Before Prompt, OpenCode Only 7k**
*We measured what Claude Code and OpenCode spend before reading your prompt, then added instruction files, MCP servers and subagents to the bill.*

---

### Introduction
When developers invoke an AI‑powered coding assistant, the visible interaction is just the tip of the iceberg. Beneath the surface, each platform consumes a substantial number of tokens to initialize its environment, load context, and prepare auxiliary services. A recent benchmark that isolated this “pre‑prompt” overhead revealed a stark contrast: Claude Code expends roughly 33 k tokens before it even sees the user’s request, whereas OpenCode stays under 7 k tokens for the same preparatory work. The discrepancy raises questions about efficiency, cost, and the trade‑offs each architecture makes to deliver its feature set.

### Key Developments
The measurement was performed by instrumenting the token counters of both systems while they executed a standard startup sequence: loading the base model, reading configuration files, initializing MCP (Model‑Context‑Protocol) servers, and spawning any declared subagents. Instruction files—custom prompts that shape the assistant’s behavior—were then added to the tally. Claude Code’s higher token consumption stems from its extensive use of layered MCP servers that maintain persistent state across sessions, plus a richer set of built‑in subagents tasked with code review, security scanning, and documentation generation. OpenCode, by contrast, adopts a leaner initialization routine, deferring many auxiliary services until they are explicitly invoked, which keeps its pre‑prompt footprint low.

### Industry Analysis
From a cost perspective, token usage translates directly into operational expense for providers that bill per‑token or per‑compute‑hour. Claude Code’s 33 k‑token overhead could add a noticeable surcharge for high‑volume users, especially when multiplied across thousands of daily sessions. However, the same investment yields immediate access to advanced capabilities—such as real‑time refactoring suggestions and automated test generation—without additional latency once the session is live. OpenCode’s leaner model may appeal to cost‑sensitive teams or those who prefer to instantiate specialized tools on