micky@garden:~/token-billing-explorer$cd ..

claude code · token economics

Tokens, turns, and prompt caching

Every turn re-sends the prior conversation. The context-window widget makes that look free; it isn't. This timeline shows the bill actually growing, and how prompt caching keeps it tractable. The simulated session below is me (Pichaya “Micky” Puttekulangkura) poking at the garden during an evening. If you want the full classroom version of this material, I teach it at Mervia Academy.

turn-by-turn billing

Watch the bill grow. Then turn caching off.

Advance one turn at a time. Each card breaks down what was billed: cache writes at 1.25× the input rate, cache reads at ~10%, new tool calls and user prompts at full price. Click Wait 5+ min before a turn to expire the cache and re-pay the write premium.

This interactive widget works best on a larger screen. The short version: every turn re-sends the full conversation, but prompt caching bills repeat tokens at ~10% the regular rate (with a 5-minute TTL). See anthropic.com/news/prompt-caching for details.
Tokens billed across turns
Every turn re-sends the prior conversation. Prompt caching bills the repeats at ~10% the regular rate.
With caching0
Without caching0
With caching
0
Without caching
0
$ claude
Advance one turn at a time and watch the billing total grow. Every turn re-sends what came before.