Prompt caching in action

Why it matters

Caching pays only when the prefix stays byte-identical across requests. An architect has to see how a small edit turns a cache read back into a cache write before relying on caching in a cost model.

What to practise

Set cache_control: {"type": "ephemeral"} on the last tool schema and on the system prompt. Send repeated requests and read cache_creation_input_tokens on the first and cache_read_input_tokens on the follow-ups. Then change one character of the system prompt and watch the tools stay cached while the system prompt is written again, because the cache follows the order tools, system, messages. Note the 5-minute default lifetime that each use refreshes, and the 1-hour option at extra cost.

How it fits

This is 2.5.1’s stable-prefix design in miniature: use the measured token counts as inputs to its break-even calculation. Then ask 2.5.4’s question of each cached block — whether its change cadence suits caching at all.