Jev + Opus 5.5: optimizing an agent loop cuts cost 10x
Across 620 requests, pinning cache breakpoints and clearing dead tool results from 214 turns cut a run from $418.50 to $41.85.
JEV + OPUS 5.5 - the same agent for $41.85 instead of $418.50. 620 requests, one model. it profiles the loop before it changes anything, pins four cache breakpoints, clears dead tool results out of 214 turns, and waits for four human clicks before it touches effort on a live route. 79.36M input tokens became 28.60M. 93% of them now read from cache at $0.50 instead of $5.00. exactly 10.000x, $376.65 off a run-day. ten times cheaper is 90% off. a 1000% cut is not a number. then I ran the same harness against a codebase that already caches. $126.40 → $41.85. 3.02x. and 62 of the 620 came back worse at medium effort and had to be re-run. the 10x was measured against a loop resending 128k of history 620 times, and the thing that made it that bad was one datetime now() in the system prompt. eleven weeks at $331 a day before anyone opened usage. one dashboard like this every few days. if this lands, someone on your team already knew it, forward it to them. how much of your 10x is actually one line nobody read?
This post quotes another post: x.com/0xspecter_/status/2104230550240174369