What is your re-billed percentage? Mine is 4.18%

On 123 sessions read 2026-09-17, 4.18% of my spend was tokens I paid for twice. That covers 105.2M tokens and $620.47 at list prices.

That is what a broken prompt cache does. The provider re-bills the whole conversation at write prices. Nothing errors. Nothing warns you. The bill just looks like you used the thing more.

Here is the part I want to be careful about. 4.18% is one machine, mine, on one reading. It is not a rate that applies to you. It moves with how you work: how long you pause between turns, how often you edit early context, how big your system prompt is, and which provider you are on, because Anthropic and OpenAI account for cache tokens differently. One machine tells you the method works. It tells you nothing about anyone else.

So the number I actually want is yours.

go install github.com/RedRobotKK/Replay/cmd/replay@latest

replay cost ~/.claude/projects

It runs locally and does not send anything anywhere. If you use OpenAI Codex, replay codex reads those logs too.

I am curious about two things: your re-billed share, and if you run replay diff, what is causing it. A re-render happens constantly and costs little. A TTL expiry is rare and expensive. On my corpus those two swapped places between readings, which is exactly why one machine is not enough.

Post your percentage and how many sessions it covers. Observations are welcome too if you would rather not run anything. I will collect what people share and publish the spread with dates and sources, same as every other number on the site.