Prompt caching: use for stable prefixes over 2k tokens; expect 30–45%, not 60%

x-token-cost-ledger instrumented six client patterns for ten weeks and attributed every token to a cache hit, a cache miss or an uncached call. Across the six patterns, prompt caching reduced input spend by between 31% and 44% (c-token-cost-reduction-1). The single-call 60% number that vendors publish was reproduced on one pattern with a 14k-token system prompt and a tight loop, and on nothing else. That is the finding behind g-prompt-caching-universal.

The reasons the number is lower in practice are mechanical, not mysterious. Cache TTLs are short (five minutes on two of three providers), so anything with human latency in the loop misses. Prefixes drift because tool definitions change per call. And a good share of spend is output tokens, which caching does not touch at all (c-token-cost-reduction-2).

The threshold of ~2k tokens is where the cache-write premium is paid back inside two reuses on current pricing. Below it, caching is neutral to slightly negative. Above it, and with the prefix pinned, the saving is reliable. The ledger is the artefact that lets a team see which of their calls are actually hitting.

Strength is moderate rather than strong because pricing moves. Two providers changed cache-write pricing between June and August, and one shortened its TTL. This recommendation carries a 90-day half-life for that reason; the number itself lives in sa-cost-per-mtok, which refreshes faster. The 'not 60%' half is the part we are confident of (c-token-cost-reduction-4).