Prompt Caching

In one sentence

Prompt caching reuses previously processed prompt prefixes so repeated large context can be billed or computed more cheaply than full reprocessing—when the product supports it and the prefix matches.

Why it matters

Agent sessions often resend large stable context (rules, docs, lockfiles). Caching can cut cost without shrinking useful context—when it actually hits.

How it works

• Stable content at the beginning of the prompt is the usual cache candidate.
• Cache hits typically require matching prefixes and provider-specific rules (TTL, minimum size, explicit cache breakpoints).
• Cached reads are often priced below uncached input; writes may have their own price.
• Volatile content should sit after the stable prefix so it does not bust the cache every turn.

Example

Put long-lived AGENTS.md and architecture docs before the changing user task and tool outputs. If the huge prefix changes every message, you pay as if caching barely exists.