Prompt Caching Explained — Save 80–90% on Repeated Context

What caching stores

Caching stores the KV (key-value) cache computed from a stable prefix — typically your system prompt, document context, or few-shot examples. On subsequent calls with the same prefix, the model skips re-processing it. Cache writes cost slightly more than standard input; cache reads cost significantly less.

Provider pricing comparison

Anthropic: write costs 1.25× (5-min) or 2.00× (1-hour) standard input; reads cost 0.10× (10× cheaper). OpenAI: no explicit write cost; cached tokens (system prompt matched) cost 0.25–0.50× standard input. Google: explicit cache write + $0.00001/token-hour storage; reads at 25–75% discount.

The TTL expiry trap

Anthropic caches expire after 5 minutes (default) or 1 hour (extended). If your calls are spaced more than 5 minutes apart, every call is a cache write — and writes cost MORE than standard input on 1-hour TTL. Check your call frequency before enabling caching.

What to put in the cache prefix

Ideal cache candidates: (1) System prompt + tool definitions (stable, large), (2) Document context for repeated queries on the same document, (3) Few-shot examples. Not ideal: variable user content, conversation history (changes every turn).

Break-even calculation

Anthropic 5-min caching breaks even at 2 calls (the second call covers the write premium). 1-hour TTL breaks even at 3 calls. OpenAI implicit caching breaks even at the first matching call. Any stable prefix used more than 3× per TTL window should be cached.