Claude Code Burns 4.7 Times More Overhead Tokens Than OpenCode

A benchmark by software firm Systima found that Anthropic's coding CLI, Claude Code, transmits roughly 4.7 times more auxiliary metadata to the underlying model than rival tool OpenCode for equivalent tasks.

The test ran Claude Code version 2.1.207 and OpenCode 1.17.18 on the same machine against the same Claude Sonnet 4.5 model. Researchers routed traffic through a proxy server to log raw input tokens, cache read/write operations, and output generation.

To measure the bare minimum operational footprint, researchers prompted both agents to simply reply "OK"—a 22-character user message. Claude Code accompanied that one-word demand with 27,300 characters of system instructions, roughly 100,000 characters defining 27 built-in tools, and another 8,000 characters of auxiliary context. That tool definitions bundle alone consumed approximately 24,000 input tokens.

By comparison, OpenCode injected about 9,300 characters of system prompt and 20,900 characters defining its ten tools, amounting to roughly 4,800 tokens for tool schemas. Stripping all tool definitions reduced the gap but did not eliminate it: Claude Code's base system prompt still consumed 6,500 tokens versus OpenCode's 2,000.

In a practical test requiring the tools to read a file and write a summary, both tools succeeded. However, Claude Code consumed around 199,000 input tokens over six API calls, whereas OpenCode finished the exact same task in four requests using roughly 41,000 tokens.

The overhead compounded further when stacking features. Adding an instructions file (a 72 KB AGENTS.md) tacked on 20,000 tokens per prompt across both clients, while five small Model Context Protocol servers added another 5,000 to 7,000 tokens. Delegating work to subagents proved the heaviest multiplier: in one test, Claude Code's token burn jumped from 121,000 tokens solo to 513,000 tokens when employing two subagents.

Source: habr

Primary source: Systima