Using different models for Plan vs Execute

[wooklab]

Hi Codex team 👋

I wanted to share a concrete example of why separating Plan vs Execute configuration could be useful.

In my workflow, an ideal setup would look like this:
Plan mode

• Model: gpt-5.2 (high reasoning)
• Used for architecture decisions, refactoring strategy, and risk analysis

Execute mode

• Model: gpt-5.2-codex-medium (or even a lower-cost model)
• Used mainly for mechanical implementation: applying diffs, editing files, wiring code

example

[plan]
model = "gpt-5.2 high"

[execute]
model = "gpt-5.2-codex-medium"

(or via CLI flags per phase).

Curious whether this kind of separation aligns with Codex’s design direction, or if similar flexibility is being considered.

Thanks!

[duguyue8]

Strongly support

[wooklab]

With the recent 0.105.0 update, it is now possible to configure the reasoning effort for Plan Mode separately.

Within a Codex CLI session, you can switch to Plan Mode using Shift + Tab, then run the /model command to select a different reasoning level. When doing so, you will see two application options:

1. Apply to Plan mode override (Always use extra high reasoning in Plan mode.)
2. Apply to global default and Plan mode override (Set the global default reasoning level and the Plan mode override.)

If you choose option 1, you can use different models (or reasoning levels) for Default Mode and Plan Mode.

After applying this setting, you can confirm that the plan_mode_reasoning_effort attribute has been added to the ~/.codex/config.toml file.

Thank you for providing a UX that allows this configuration directly within the session.