64 个 agent 前置 Jev 分流：Claude Code 月账单从 4.1 万降到 7800 美元
苏黎世工程师让 Jev 先处理每天 14 亿行里的简单判断，只把置信度低于 0.91 的 17% 交给 Opus，并用夜间校验器动态上调阈值。
this f*cking thing just cut a Claude Code bill by 81% a 24-year-old ML engineer in Zurich gave Claude Code a second brain with Jev. 1.4 billion rows a day now go through it, and the Opus bill fell from $41,000 a month to $7,800 his 64 agents used to send every tiny yes/no to Opus. is this email spam? which agent takes this ticket? is this row a refund? so he put Jev at the door. the easy rows it settles on the spot, and only the ones it can't call go upstairs to Opus before: every row -> Opus -> $41,000 now: every row -> Jev -> the hard 17% -> Opus -> $7,800 Jev scores how sure it is on every call, and that score decides who goes upstairs. he didn't guess the cutoff. the 20-minute test, on your own data: - run 200 real rows through Jev and Opus both - sort Jev's answers by how sure it was - find the score where Jev stops disagreeing with Opus. his was 0.91 - above it Jev's answer is final, below it the row goes upstairs a verifier sits on top of the swarm. every night it sends 50 of Jev's most confident answers to Opus anyway, and two disagreements in a week push the cutoff up. refunds, deletions and anything sent to a person never take a Jev-only answer works anywhere the question is a label: support tickets by urgency, leads hot or cold, comments fine or spam, a category column across 40,000 rows Opus still reads every hard row. it stopped reading 1.16 billion easy ones a day
这条是引用着另一条写的：x.com/thegreatest_sv/status/2102802378378621355