按官方价算 Jev 与 Opus 5.5 的 agent 调用成本

werden @0xwerden原推发布于 2026.10.03

在 X 上打开原推

Everyone's posting jev + opus 5.5 savings. Almost nobody shows the math. so I ran it on list prices. say your agent makes 10,000 small calls a day. which file to open, retry or stop, which test first. each call sends ~2,000 tokens of context and gets a short answer back. all on opus 5.5 ($2 in / $10 out per 1M): -> 20M input tokens = $40 -> 500K output tokens = $5 -> about $45 a day, before caching the same calls on jev ($0.042 per 1M in, output free): -> 20M input tokens = $0.84 -> that's the whole bill even if 1 in 10 calls is unsure and still goes up to opus: about $5.34 a day. opus should be paid to think, not to choose between three options. what the math doesn't tell you: -> jev's accuracy numbers come from typesafe's own benchmark -> a wrong cheap pick that opus has to clean up isn't cheap -> caching cuts the opus side a lot when your context repeats so don't move everything. move the calls with a short answer list and a check you can run. how many of your agent's calls are really just picking from a list?

这条是引用着另一条写的：x.com/polydao/status/2103689373774483815