Opus 5.5 规划 + Jev 路由 + GPT-6.1 Sol 执行的成本推演

mojesko @mojeskoqq原推发布于 2026.09.30

在 X 上打开原推

This is the first AI setup whose math genuinely shocked me opus 5.5 plans. jev decides. gpt-6.1 sol does the work. using devday's per-task numbers, on paper the three together cost about a quarter of opus alone alone, each one hits a wall. jev can't write a single line. 6.1 sol is cheap, but i'd still hand opus 5.5 the long open-ended work, and it bills $4 in / $20 out per 1M tokens on every tiny step i haven't seen anyone run all three together. so i mapped them into one loop, step by step, at list prices: → one loop, three models, 8 step types. every step has exactly one owner. the video is the sim log → opus 5.5 plans the run. it touches the plan and the escalations, nothing else → jev reads every step after the plan and answers one question: opus, sol, or plain code. $0.042 per 1M tokens in, output free, 70-500 ms → one $0.12 opus plan call buys 2,857 jev decisions at 1,000 tokens each. that's the balance at the end of the video → 6.1 sol writes, reviews and fixes. on terminal-bench science at max effort it runs $5.47 a task. opus 5.5: $23.21 on the same benchmark → sol re-reads the repo from cache. cached input is $0.10 per 1M, so re-reading a 200K-token repo costs $0.02 on paper, a $23.21 opus-only task becomes about $5.59: sol's $5.47, one $0.12 opus plan, $0.00042 for 10 jev calls. cost only, i'm not claiming it scores like opus that's the stone in the video. the rough is priced like opus alone. every cut is money not spent a lot of stacks pay their smartest model $20 per 1M output tokens just to say "skip". this one pays jev $0.042 per 1M in and nothing for the answer on sep 29 openai put its own decisions api into limited preview. the money is moving into the router layer every day you wait, you pay opus prices for work that costs about a quarter on paper bookmark this before your next opus invoice lands

这条是引用着另一条写的：「一篇上手 Jev 的完整构建指南：代码、架构与图」