Claude Code sends 33k tokens before it even reads your prompt; OpenCode uses 7k
A blog compared token usage between Claude Code and OpenCode on identical tasks. The trigger was a simple observation: after switching from OpenCode to Claude Code, the usage counter climbed noticeably faster.
To check the hunch, the authors dropped logging between the agent and Anthropic's endpoint and captured every request along with its usage blocks. Their numbers: Claude Code fires off around 33k tokens before it reads the actual prompt, while OpenCode gets by on roughly 7k.
They pin the gap on Claude Code's less efficient caching strategy and the harness's heavier overhead of housekeeping tokens. There's one caveat they add near the end of the post. The Hacker News thread drew 46 points and 16 comments.
Related stories
- Auditing one Claude session: 7.5M tokens, 4,038 tool calls
- By the end of a long session, strict solutions drop to 0.5%
- rtk promised 60-90% fewer tokens. JetBrains measured 7.6% more spend
- Claude Code implements automatic prompt caching, reducing input token costs to 10%
- Claude Code matches Codex on SWE-Bench Pro at twice the cost
- A four-model Claude ensemble scored 78% on Terminal-Bench 2.1, at double the cost of first place
Comments
No comments yet. Be the first.
Join the conversation
Sign in with Google to leave a comment. Your name and avatar come from your Google profile, and the comment appears after moderation.