26/08/2026
Nobody's AI bill is high because the model is expensive.
It's high because of five things the interface never shows you.
The session you never cleared.
Claude Code resends the entire conversation on every turn. Not a running tab - a full re-send. Finished work keeps getting re-transmitted while you work on something unrelated, and every turn is bigger than the last.
Tool schemas you forgot were on.
Every connected MCP server loads its full tool definitions into every message before you type a word. Three or four servers is 10,000-18,000 tokens per message. This is the invisible one.
A cache that quietly stopped working.
Prompt caching is a prefix match - one changed byte invalidates everything after it. Cache reads cost a tenth of normal input, so a broken cache is roughly ten times the price on the whole cached span. A timestamp in your system prompt will do it.
Reading the whole repo to change three files.
Including only the relevant files uses 60-80% fewer tokens at the same output quality. And everything you pull in gets re-sent for the rest of the session.
Top model, top effort, trivial task.
The spread between tiers is five to one. Counterintuitively, higher effort sometimes costs less on agentic work - better planning up front means fewer turns.
Number three is the one worth checking today, and it takes about thirty seconds. One field tells you: cache_read_input_tokens on a repeated call. Zero means something is invalidating it, and you have been paying ten times over without knowing.
Swipe for the detail on each.
Vibe coded yourself into a corner? That is fixable.
[email protected]
