prompt caching: one write, many reads
Request 1
[system + tools + few-shot] 15,000 tok (CACHE WRITE)
[user msg] 400 tok
bill: 15,000 * 1.25x + 400 * 1x
Request 2
[same prefix] 15,000 tok (CACHE READ)
[user msg] 400 tok
bill: 15,000 * 0.1x + 400 * 1x
Request N
[same prefix] 15,000 tok (CACHE READ)
[user msg] 400 tok
bill: 15,000 * 0.1x + 400 * 1x
Anti-pattern: timestamp at the top
"15:30:02"
[rest of prefix] still 15,000 tok
[user msg]
every request MISSES: no shared prefix
break-even: Anthropic pays back after the 2nd read; 1h TTL costs 2x write but reuses 12x longer.