Long Claude Code sessions cost 31× more per request, logs show

The post is a practitioner's cost audit rather than a vendor announcement. The author, writing as iguchi of Mukuil, a startup building AI solutions for manufacturing, first assumed that because cache reads are cheaper per token, caching was keeping spend down. Aggregating the usage logs showed the opposite: cheap tokens were being re-read at enormous volume, and that volume was the bill.

The mechanics explain why. Claude Code sends a request to the API on every tool call, and each request carries the whole conversation so far — system prompt, CLAUDE.md, files read, tool output. The arithmetic in the post is that if context grows by a roughly fixed amount per request, total session input scales with the square of the request count, so the back half of a long session is where cost concentrates.

The team's fixes aim at both factors in that multiplication. Session hygiene (/clear, /compact with specified content, restarting past 300K tokens) shortens context, while batching independent tool calls reduces request count; the post estimates moving from an average of 1.17 to 2 calls per message would cut requests by about 41%. One detail suggests the organizational side is the harder part: a subagent defined for large-file reading went unused for 97 sessions and 21,095 requests until CLAUDE.md stated it as a standing instruction, after which it was used 11 times in 27 sessions.