Codex tip: a cost-effective GPT-6 Luna + GPT-6.1 Sol (medium) agent tree, orchestrated by GPT‑6.1 Sol at low effort.
6.1 Sol replaces Astra at 1/5th the cost.
give your Luna the narrow work, and call Astra only when an independent review is needed.
use Terminal Bench's pass rates and cache costs as a starting signal, then tune the effort levels to your own work.
hand this to Codex to set it up 👇