Quick answer
A cache write token is context the provider stores in the prompt cache so later turns can read it cheaply. Writes are priced above normal input, so a cache write only pays off if the prefix is reused. Editing the prefix repeatedly turns every turn into a cache write, which is the worst case.
When a write pays off
| Scenario | Writes | Reads | Result |
|---|---|---|---|
| Long session, stable prefix | One | Many | Cheapest |
| Short one-off task | One | None | Slight premium |
| Prefix edited each turn | Every turn | None | Most expensive |
| New session per follow-up | Every session | None | Wasteful |
Avoid wasted writes
- Do not edit AGENTS.md or CLAUDE.md mid-task.
- Do not prepend new instructions before existing context.
- Keep the same model for a session; caches do not cross models.
- Continue the session for related follow-ups.
Verify the balance
Compare cache write and cache read totals over a session. Writes should be a small fraction of reads in a long conversation. If they are close, the prefix is churning.
The cache read guide explains the other half of the pair.
Cache writes are an investment
Storing a prefix for reuse costs more than sending it once, and pays back only if you send it again. A short session that ends early leaves you with the surcharge and no discount, which is why cache writes appear on bills for work that felt small.
- Cache writes make sense for repeated long context.
- One-off requests should not pay to store a prefix.
- Keep instructions stable so the stored prefix stays valid.
- Read the cache read guide for the payoff side.
UsageAtlas separates the two fields in local history, so you can see whether caching is helping or costing on a given project.
Questions, answered
Are cache write tokens more expensive than input?
Usually yes. That premium is the cost of storing the prefix for later reuse.
Why do I see cache writes on every turn?
The prefix is changing. Something early in the request, such as instructions or a system message, is being edited.
Do cache writes count against my plan limit?
Yes. They are usage like any other token, even though they are reported separately.