Routing with GPT-6 Astra doubles useful work on the same Codex limits
The author splits routine agent work to cheaper models, keeping Astra on direction and Sol on implementation, and says this doubles output under the same Codex quota.
THE $200 CODEX PLAN SHOULD NOT SPEND ITS BEST MODEL ON ROUTINE AGENT WORK. i moved the cheap loops out of it. the stack: → GPT-6 Astra directs the project → DeepSeek V4.1 Flash handles discovery + routine checks through $10/month OpenCode Go → GPT-5.6 Sol owns implementation → GPT-5.6 Luna reviews only when the risk justifies another call that doubled the useful work i get from the same Codex limits. setup takes three moves: → connect OpenCode Go to Codex through model-router → install Quota Flow with its skill + agent profiles → start a fresh Astra task and tell it how to route the work the important detail: the same Sol agent keeps its context through: build → test → fix no new worker relearning the repo at every step. the breakthrough isn’t adding more agents. it’s keeping Astra on direction, cheap models on volume and context with the agent already doing the work. route by responsibility. make every additional agent earn the context it consumes. save this setup.
This post quotes another post: x.com/av1dlive/status/2098055179761525165