If you use Claude Code heavily, your token bill may be inflated not by your prompts, but by the tool's own scaffolding. A small empirical study by the Systima team found that Claude Code transmits roughly 33,000 tokens to Anthropic's API before it even processes your first instruction. OpenCode, an open-source alternative that routes to the same Claude models, sends around 7,000 tokens to accomplish the same setup — about 4-5x less overhead per session.
The team arrived at this finding by instrumenting a proxy layer between each coding agent and Anthropic's endpoint, logging every request and the associated usage data returned in response blocks. The trigger was practical: they'd been forced off OpenCode temporarily and noticed their usage meter climbing noticeably faster with Claude Code. The logging study converted that anecdote into hard numbers.
The gap comes down to two factors: cache strategy and harness design. Claude Code's system prompt and tool definitions are larger and appear to be structured in ways that limit cache reuse, meaning more tokens get re-sent on each turn. OpenCode's leaner harness sends less redundant context. The difference compounds across a multi-turn coding session — those extra 26k tokens fire on every interaction, not just once.
Why it matters: if you're paying for API access directly (or your org is), this overhead is pure cost with no quality return. Both tools are calling the same underlying model. The intelligence is identical; the wrapper efficiency is not. For teams running dozens of sessions daily, the cumulative token waste is substantial.
What you can do: if you're on a token budget or optimizing for cost, OpenCode is worth evaluating as a drop-in alternative. It supports the same Anthropic models and, based on this data, runs significantly leaner. The Systima post notes one caveat they encountered — worth reading in full before switching production workflows — but the directional finding is clear: tool choice affects your token spend independently of model choice.