Jev pre-filtering cuts email agent cost by 10x
For the same 10,000 emails, Jev judges first and sends only uncertain cases to Kimi K3, lowering the monthly bill from $300 to $9.42.
$300 a month vs $9.42 for the same 10,000 emails Same inbox, same decisions, different routing Most of an agent bill is not the work It's the model deciding what to do next, billed as a full frontier generation every time The fix in this breakdown: -> Frontier model reads everything: $300 -> Kimi K3 reads everything: $90 -> Jev first, K3 only on what Jev doubts: $9.42 The part that makes it work is not the price Jev returns a probability with every answer, so your code knows which ones to trust and which to escalate One caveat: -> The $9.42 assumes 10% get escalated -> In the real fraud test in the article it was 31% -> That lands closer to $28 Still 10x cheaper You don't need a smarter model for small decisions You need one that knows when it's guessing
This post quotes another post: x.com/polydao/status/2104783226833186920