5.6-Luna low consuming more usage than 5.6-Sol Low on the exact same task

What version of Codex CLI is running?

0.144.1

What subscription do you have?

Plus

Which model were you using?

5.6-Luna

What platform is your computer?

Linux

What terminal emulator and version are you using (if applicable)?

kitty + tmux

Codex doctor report

What issue are you seeing?

I decided to try Luna to orchastrate a loop, but when I was priming its context (4 prompts), 5hr my usage droped from 100% to 77%, I though this was bad, then tried a new session this time with Sol, the same exact prompts and the usage only dropped from 77% to 71%. There's something really off with Luna usage, both models on low.
Token usage for luna on the task: Token usage: total=41,510 input=38,238 (+ 384,000 cached) output=3,272 (reasoning 820).

Token usage for Sol on the same task: Token usage: total=50,296 input=46,679 (+ 347,392 cached) output=3,617 (reasoning 582)

What steps can reproduce the bug?

Basically read codex memory, read some orchestration scripts I have and set some things straight.
Prompt 1:
"ok, here is what I want you to work at: you're gonna be a loop supervisor, you need to make sure the loop keeps running, you'll check from time to time how the loop is doing. you have some instructions in codex memory and scripts at ~/github/codex-scripts. check and let me know what you learned"

Prompt 2:
"what I need you todo: 1. every 15 minutes check if the loop is healthy (waiting for prompt or stuck on codex messages or out of usage, communication, problems, agent waiting for another agent reply which isn't replying, etc) and handle any problems by resuming agents, asking them to continue, resuming goals, switching codex accounts. you ALWAYS need to usa an external script to wake you up. you have the regular 30 minutes, just change it to 15 minutes. I'll prompt the other orchastrator on how the loops works, you just need to make sure they don't stop unless the work was actually deemed as done. got it ? kill any timer left from old loops/ supervisor"

Prompt 3:
" ok, but when you wake, you check not only orch2, but also the agent in charge of the current task to see if it's doing well"

Prompt 4:
" also that, in the eventuality you need to switch sessions accounts, to the set correct model for each session on restart: builder gets gpt-5.6-Sol low ; codex-orch gets gpt-5.6-luna low ; both reviewers get gpt-5.6-Sol high when you switch accounts codex uses the last model select, so when restart sessions this can cause a model mismatch from the intended loop, be mindful about it."

that's was basically a fairly "read codex memory and some scripts" + some followup instructions on how to handle the task, and 23% gone.

What is the expected behavior?

_No response_

Additional information

_No response_