model_reasoning_effort

model_reasoning_effort rides on top of whatever model you run and tells it how much thinking to spend per turn. The official reference gives it one line: “Adjust reasoning effort for supported models (Responses API only; xhigh is model-dependent).” That’s the whole entry (most of what’s actually known about this key lives in the issue tracker).

Values

Key	Values	Default
model_reasoning_effort	minimal, low, medium, high, xhigh	unset (model’s own default)
plan_mode_reasoning_effort	none, minimal, low, medium, high, xhigh	unset (plan mode’s built-in preset)

The docs don’t publish what “unset” resolves to per model. If the number matters to you, set it explicitly:

# ~/.codex/config.toml
model_reasoning_effort = "high"

# plan mode keeps its own preset unless you override it
plan_mode_reasoning_effort = "xhigh"

When to change it

1. Limits draining fast. The opening report in “Burning tokens very fast” (#14593) burned roughly 20% of a business account’s tokens in two hours running high reasoning effort in VS Code. Dropping to medium or low is the first lever to pull.
2. Codex keeps whiffing on multi-file work. Go high, then try xhigh if your model accepts it.
3. Quick edits. Renames, config tweaks, and one-file fixes run fine on low or minimal.
4. Smart plans, cheap execution. Leave the main key alone and raise plan_mode_reasoning_effort. Daniel Vaughan’s delegation-spectrum writeup pairs reasoning effort with approval_policy when dialing in autonomy: the more rope you hand Codex, the more thinking you want behind each move.

Gotchas

xhigh is model-dependent and the reference doesn’t list which models accept it, so expect to test.

Higher effort doesn’t always buy more thinking. #30364 tracks GPT-5.5 reasoning tokens clustering at exactly 516/1034/1552 per response, and a related report (#29353) describes xhigh runs short-circuiting at 516 reasoning tokens. #30364 is still open with no maintainer response, and #29353 was closed as not planned, so on GPT-5.5 treat this setting as a request the model can undercut.

The docs scope the key to the Responses API only; whether it silently no-ops on other providers is undocumented.

Related settings

• model picks the model the effort applies to; xhigh support depends on it.
• approval_policy is the other half of the autonomy dial in Vaughan’s pairing.
• If you landed here because your quota vanished, see usage limit draining too fast.