What a cold cache costs in Claude Code
Published 29/10/2026
Claude Code keeps your conversation in a prompt cache so each turn re-reads it cheaply. When that cache is lost, the next turn has to write the whole conversation back in, and that one turn can cost as much as twenty ordinary ones.
Model a5v3, last checked against a live meter on 29/10/2026, on a Max 20x plan. More: plan usage burn data.
Warm turns and cold turns#
| One Opus turn, 2,000 tokens out | Share of a week | Turns like it in a week |
|---|---|---|
| 200k context, warm cache | 0.004% | about 24,000 |
| 200k context, cold cache (rewritten) | 0.046% | about 2,200 |
| 800k context, warm cache | 0.008% | about 13,000 |
| 800k context, cold cache (rewritten) | 0.176% | about 570 |
Think of a library desk. Keeping your books open costs a little rent. Putting them all back and fetching them again costs an afternoon, and at 800k context it is a lot of books: a cold turn there is about 23× a warm one.
What loses the cache#
- A pause longer than the cache lasts: an hour for a plan's main conversation, five minutes for subagents and API keys. (Setting the TTL.)
- Compaction, because the history is rewritten.
- Switching model, because each model has its own cache.
- Adding or removing tools mid-session, such as connecting a server whose tools load up front.
- Upgrading Claude Code, for the first session after.
Editing files, changing permission mode and invoking skills keep the cache.
How to have fewer cold turns#
- Pick the model at the start of a session, and stick with it.
- Compact at a break you choose, not when the window forces it.
- Keep the context lean, so the cold turns you can't avoid are small ones.
- Start helpers with a short briefing, not a pasted history.
Where ADE fits#
ADE's briefings are short by design, at most 250 tokens for the first-edit briefing and a fixed budget for the handover, so a fresh session or a builder in its own lane starts small and its cold turn stays cheap. Long command output goes to a local store (ade q run) instead of the context. See How it flies.