Guides
How prompt caching changes session cost
Caching can reduce repeated-input cost, but only under each provider’s exact rules.
A concise editorial guide, not a universal recommendation or guarantee.
Decision rule
Long agent sessions repeatedly send instructions, tool definitions, and history. A cache hit may price that repeated prefix differently, while writes, expiry, minimum lengths, or changed prefixes can remove the saving. Track cache-read and cache-write usage separately and treat missing catalogue prices as unavailable, never free.
Practical checklist
- Keep stable prefixes actually stable.
- Inspect cache reads and writes separately.
Sources checked
These first-party or authoritative sources ground the note. Checked 2026-08-16.