I want Codex Ultra to waste less budget, but I can’t add another complex division-of-labor system just to save tokens.

I just ran a minimal experiment: the main model is still GPT-5.6 Sol Ultra—it only handles complex reasoning, task decomposition, and final acceptance/review. Researching for sources, scanning code, and similar work gets delegated to the faster and more token-efficient Terra Medium, which is read-only throughout. Tasks with clear boundaries and well-defined execution paths are sent directly to Sol Medium; complex, ambiguous, and high-risk parts are kept on Ultra.
Sol Medium is still Sol, but its reasoning strength is reduced from Ultra to Medium. The official positioning of Medium is the balanced default for most tasks; pushing the reasoning strength higher may improve quality on complex tasks, but it will increase latency and token usage. So I can’t say quality is completely unaffected—I can only control risk through clear task boundaries plus Ultra’s final acceptance.
I’m still running it on real tasks right now, so I haven’t measured the specific savings rate yet, and I haven’t systematically compared quality differences. I’m observing.
Have any of you done a similar setup where a stronger model only does judgment and acceptance, while Medium handles only clear execution? Did the quality drop noticeably?