Codex CLI reasoning effort and models

There are three ways to pick the model, from most temporary to most permanent:

# For one run
codex -m gpt-6.1-sol
codex exec -m gpt-6.1-sol "explain this stack trace"

# Mid-session, inside the TUI
/model

# As your default, in ~/.codex/config.toml
model = "gpt-6.1-sol"

The model id above is the example OpenAI's config reference uses. Use whatever ids your own login lists (next section). /model opens a picker with the models available to your account and lets you choose the reasoning effort in the same step, when the model has one. Run /status afterwards to confirm what the session is now using.

-c overrides any config key for one run, so codex -c model=gpt-6.1-sol is the same as -m. Flags and -c win over config.toml.

Which models you can use depends on how you signed in. A ChatGPT login and an API key reach different lists, and so do different plans. Check which one you are on first:

codex login status
# Logged in using ChatGPT

Then print the catalog Codex has for that login. codex debug models renders it as JSON, and it costs nothing because no model is called:

codex debug models | jq -r '.models[]
  | select(.visibility == "list")
  | "\(.slug)  default=\(.default_reasoning_level)  levels=\([.supported_reasoning_levels[].effort] | join(","))"'

Each entry carries the model's slug (the id you pass to -m), its default_reasoning_level, the supported_reasoning_levels with a one-line description of each, and its context_window. Codex caches the same catalog in ~/.codex/models_cache.json. codex debug models --bundled skips the refresh and shows only the catalog built into your binary, which is useful for telling "my account cannot see it" apart from "my CLI is too old to know it".

If Codex answers a model with is not supported when using Codex with a ChatGPT account, the name is fine and your login does not cover it. See Codex model not supported on this account.

Reasoning effort is how much the model thinks before it answers. More effort means slower turns and more tokens, and better results on problems that need several steps of reasoning. Set it in any of these places:

# Default, in ~/.codex/config.toml
model_reasoning_effort = "high"

# One run
codex -c model_reasoning_effort=high

# Mid-session: /model, then pick the level

The levels belong to the model, not to Codex. OpenAI's config reference lists low, medium, high, xhigh, max and ultra as examples and says the available levels depend on the model. On the ChatGPT login we checked in October 2026, the catalog described them like this:

Not every model offered every level: some stopped at max. Older guides mention minimal, which none of those models listed. Run the catalog command above to see the levels for your models today.

The default is per model

If you never set model_reasoning_effort, Codex uses the model's own default_reasoning_level, and that is not the same everywhere. In the same catalog some models defaulted to low and others to medium. So switching models without setting effort can quietly change how hard Codex thinks. Set the key explicitly if you compare models or care about cost.

Plan mode has its own setting

plan_mode_reasoning_effort sets the level used while Codex is in plan mode (/plan). When it is unset, plan mode uses its built-in preset default rather than following your normal setting.

Which level to use

• Low or medium for edits you could describe in a sentence, renames, test scaffolding and questions about the code.
• High for debugging, refactors across several files and anything where a wrong first guess is expensive.
• The top levels for a stuck problem, not as a default. They spend your usage limit fastest, and on a ChatGPT plan that is the budget that runs out. See Codex usage limit reached.

Two more keys shape what you see rather than how hard the model thinks:

model_reasoning_summary = "concise"  # auto, concise, detailed, none
model_verbosity         = "low"      # low, medium, high

• model_reasoning_summary controls the short summaries of the model's reasoning that Codex shows while it works. none turns them off. A related boolean, model_supports_reasoning_summaries, forces Codex to send or not send reasoning metadata, which matters for a model or provider Codex does not describe correctly.
• model_verbosity controls how long the model's written answers are. When unset, the model's default applies. The catalog says per model whether it supports verbosity at all.

For speed, service_tier = "fast" (or /fast in the TUI) selects a faster tier when the model offers one. The catalog we checked described it as 2x speed with increased usage.

If you move between a cheap everyday setup and a deep one, put each in a profile. In the current CLI a profile is a file next to config.toml named <name>.config.toml, layered on top of your base config with -p:

# ~/.codex/deep.config.toml
model_reasoning_effort  = "xhigh"
model_reasoning_summary = "detailed"

codex -p deep
codex exec -p deep "find the race in the job queue"

Codex ignores profiles set in a project-local .codex/config.toml, so keep them in your user config. Profiles can carry sandbox and approval settings too, covered in Codex sandbox and approval modes.

Each catalog entry has a context_window and an effective_context_window_percent. On the models we checked that percent was 95, which is where the usable figures in the cards above come from: 95 percent of 272,000 is 258,400. Codex keeps the rest for its own harness.

Two keys override this when you run a model the catalog does not describe well, such as one from a custom provider: model_context_window sets the window, and model_auto_compact_token_limit sets the token count at which Codex compacts the history on its own. /compact compacts on demand.

Termdeck's model picker reads the Codex catalog on the machine rather than a list it ships, so a new model or a new effort level shows up with whatever levels it comes with. A chat left on the default runs at the model_reasoning_effort from that machine's ~/.codex/config.toml, and the picker names it. Picking a level pins it for that chat, even if your config changes later. A change of effort applies from the next turn, not to a turn already running.

Because the setting is per chat, you can keep a high effort chat on a hard bug and a low effort chat on chores side by side, on one machine or several. See Engines for the per engine differences, and Claude Code models and effort for how the same dial works there.