Keep Sol on the decisions. Send small, checkable jobs to Luna Max. One prompt sets it up.
A stronger model should handle the parts where context and judgment matter. A cheaper one can handle small jobs with a clear scope and a check at the end.
01
Sol
Orchestrates and decides.
02
Luna Max
Implements the bounded job.
03
A check
Proves the job is done.
Why bother
Luna Max lands near Sol Medium on the measured score at a lower cost per task. That is one benchmark, not a general equivalence.
Create the worker
Paste this into Codex. It makes Codex inspect the installed setup, keep the rest of the config, show the diff, check it against the installed version, and only then start using the worker.
codex-luna-worker.txt
create a custom agent named luna_worker at ~/.codex/agents/luna-worker.toml.
use these settings:
model = "gpt-5.6-luna"
model_reasoning_effort = "max"
give it a description and instructions for bounded delegated work. preserve the rest of my config. validate it against my installed Codex version, show me the diff, then use luna_worker for subagent tasks.
What goes where
Keep with Sol
Judgment
Architecture and trade-offs
Ambiguous or shifting requirements
Security and high-impact changes
Send to Luna Max
Throughput
One named function, component, migration or test
A clear file or folder boundary
A check that proves it is done
Want this in your stack?
I set up agent workflows that are cheap to run and easy to trust.
