Context is the bill
Why a 4,000-turn session costs more than forty 100-turn sessions, and what to do about it.
4 min
Every turn resends everything
Claude Code has no memory between API calls. Each turn sends the whole conversation again: system prompt, every file you read, every tool result, every reply. Caching reprices that resend to about a tenth of list price, but a tenth of 900k tokens is still 90k tokens per turn.
That is why the biggest number in most ReadyClaw reports is context bloat, not model choice. A session that grows to 500k tokens and runs another 2,000 turns pays for 500k tokens two thousand times.
What to do
Start a fresh session per task. When you finish a feature, write a five-line handoff (what changed, what is verified, what is next) and open a new session with it. The new session starts at 10k tokens instead of 500k.
Run /compact when a long investigation is over and the details no longer matter. Compaction keeps the conclusion and drops the transcript.
Prefer the 200k-context model variant as your default. The 1M variant is for the rare task that genuinely needs a huge working set; as a default it simply removes the ceiling that used to force you to compact.
npx readyclaw scan