A turn at 141 costs 2.1x a turn at 20 in Claude Code

In his own Claude sessions, any turn after turn 141 cost 2.1x the average turn in that session’s first twenty, says Tom, the developer behind the public task tracker my-status.app. He counted 144 sessions and 14,640 turns, and he published the curve and the script that produced it in a September 3 Show HN post on Ycombinator’s Hacker News.

The idea came from Claude. On August 26, Tom asked in a CoWork session what else they could do with what they had learned about how sessions work. Claude came back with six ideas. One was “the session aging curve (when a fresh seat is cheaper)”. Another was “the scaffolding trend (the cost of our own rules, creeping 87,999 → 92,619 → 93,809 across three seats)”. The numbers came four days later.

The mechanism behind the curve is simple. The model keeps nothing between turns, so every request carries the whole conversation so far. Prompt caching bills that re-sent history at 10% of the input price. The post measures this standing context, which is re-read on every turn, at 114,133 tokens on the “hub” seat and 108,213 on a “spoke”.

The counting went wrong on its first run. The script reported 46.6M tokens where the real figure was 28.1M, so it overcounted by about two thirds.

The post also tracks the human’s share of the wall clock, which comes to 67%. Every feature from now on comes with a receipt, Tom writes. The aging curve took 2.6 person hours and $8.23 in tokens at list rates.

His own batch sessions were the worst offenders, so he stopped batching. He describes his setup as his own version of “one thinking session, short task sessions”, meant to make better use of his plan. He had kept running out of Fable usage while about half of his usage for Opus was left over.