Animated cost comparison explaining prompt caching behavior

the same 50-call agent run on opus 5.5 costs $0.74 or $12.50. the difference is one timestamp in the system prompt a 50k-token prompt, resent on every call: > no caching: 50 × $0.20 = $10.00 > caching on: one $0.25 write, then 49 reads at $0.01 = $0.74 > caching on, "current time" at the start of the prompt: every call misses and writes again, 50 × $0.25 = $12.50 the cache matches from the first token. change anything near the start and the whole prompt is billed as new, at 1.25x the input price catch: the cache lives 5 minutes, so an agent that idles longer pays the write again check cache_read_input_tokens on your second call. 0 means something near the start moves. the other 6 settings are in the article

— 0xbobaa (@0xbobaaa) Sep 29, 2026