Study: Claude Code Burns Through Up to 33,000 Tokens Before It Even Reads the User's Request

A post comparing the token overhead of two AI-based development tools — Claude Code and OpenCode — has sparked a wide-ranging discussion on the Hacker News aggregator.

According to the figures presented, Claude Code sends the model on the order of 33,000 tokens before it even begins to process the request the user actually typed in. By comparison, OpenCode makes do with roughly 7,000 tokens in the same situation.

The post racked up 265 points and set off a lively debate — 140 comments have already been left under the material. This level of interest from the community shows that the topic of token-usage efficiency in developer AI tools remains a sensitive one: the volume of housekeeping data passed to the model with every request directly affects both response speed and the cost of using such services.

The authors of the study do not disclose the full measurement methodology in the available description, yet the mere fact of such a significant gap — more than fourfold — between two tools tackling similar tasks raises the question of the architectural decisions underlying the system prompts and the context these tools assemble before handing the request off to the language model.

The discussion on Hacker News suggests that questions of optimizing overhead are becoming ever more important as the number of users working with AI coding assistants day in and day out grows. The more tokens go toward housekeeping information, the higher the ultimate cost of the work and the more noticeable the delay before an answer comes back.

Source: Hacker News