Jev 做路由：Flash 干活、Opus 5.5 只读 diff 的成本分层实验

yurshev @yurshevv原推发布于 2026.10.22

在 X 上打开原推

STOP PAYING OPUS TO F*CKING COSPLAY AS THREE MODELS. one guy. three name tags. one invoice shaped like a team of five. that is your "agent swarm" if nobody set the routing layer. CLAUDE CODE / AGENT LAB fixes it: Jev decides. Flash works. Opus 5.5 only reads the diff. one pipeline, three price tiers, three jobs that are not interchangeable. the quiet leak you built yourself: Opus burns $20/1M output on files it had no business reading. Flash sits idle. every subagent silently defaults to the most expensive model. you discover it on the invoice, not on the dashboard. the four laws printed on the panel: → VERIFY CLAIM: p(source) ≥ 0.90 or the claim gets parked, not shipped → COST GATE: Flash caps at ~$0.20/task before the router escalates → RETRY RULE: tries ≤ 2, then the branch parks for human review → REVIEW: Opus 5.5 reads diffs only, never raw files one honest line: this is an illustrative lab, not a benchmark. the point is the shape, not the exact dollar. a one-shot script doesn't need three tiers. a 2-hour agent run absolutely does. nobody wants to admit how much Opus they waste on grep.

这条是引用着另一条写的：x.com/bober_smart/status/2078784709253841039