Codex effort guides how much reasoning the model applies to a task. Higher reasoning effort can help with complex work and typically increases latency and token use, though the result depends on the task and the model. In a reply on GitHub on December 25, 2025, user etraut-openai recommended medium for most coding tasks and low for simple ones. That is a useful starting point, not a rule for every model. Raise the level only when a task fails at the level you run.
How do I decide which level of reasoning to use in Codex?
A user on the OpenAI developer forum asked this in February 2026, while planning to have Codex specialize many theorems in one batch. They wanted to know whether the level sets the total reasoning after a prompt, or how carefully each subtask is done. One reply suggested testing every level and keeping the lowest one that works. Another community member described the level as a cap on the reasoning tokens the model may spend before it answers (the forum thread). That is one member's reading, not an official definition.
The official definition is the better guide. Reasoning effort guides how much reasoning the model applies to a task, and the docs do not publish a token count for each level (OpenAI reasoning guide). A token limit is a separate, explicit setting: in the API, max_output_tokens caps the tokens a response generates, reasoning included. The forum's first reply still gives a workable test. Run a task whose right answer you already know at two levels, compare the results and the time, and keep the lowest level that gets the answer right.
- Choose one task whose correct result you already know.
- Run it at medium, then at high, and note the time and whether the result is right.
- Keep the lowest level that produces the right result.
- Repeat the check when you switch models, since the levels on offer differ by model.
The levels a session can use depend on the model and on the client. The Codex models page says the options and defaults vary by plan, client, workspace settings and rollout, so the picker in your own app is the list that counts (Codex models).
Can Codex low match medium on easy tasks?
The forum thread asks this question but reports no side-by-side test. In a thread opened on February 1, 2026, a user who ran medium for simple work and high by default, to manage a weekly usage limit, asked whether low could take some of the simple tasks. The poster's impression was that medium and high often perform alike on work that is not very challenging. One reply agreed that a well-defined simple task, or one backed by a strong skill file, can work at low. It also quoted a Codex commit that set plan mode to medium after high was reported to overthink and run past the context window (the forum thread).
The official wording gives the same trade-off: higher effort "can improve results for complex tasks, but it takes longer and uses more tokens" (Codex models). On a rename or a one-line fix, extra time and tokens are unlikely to buy much, so the useful question is whether low gets the same result on your task. The thread gives experience, not measurements. Test low against medium on your own simple tasks before you move a workflow to low for good.
Why is Codex so slow on simple tasks?
Often the level is set too high for the task. A slow connection adds to the wait at any level, and an older CLI can still run a feature that newer versions removed. GitHub issue #8545 shows both. Opened on December 25, 2025, it reports a Windows user at the xhigh level, whose git status check and push were "thinking for 50 minutes". In a reply that day, GitHub user etraut-openai wrote that xhigh is relatively slow and recommended medium for most coding tasks, high for complex work such as major refactors, and low for simple tasks. In a second reply later that day, the same user noted that a slow internet connection adds to the wait, and pointed to an experimental undo feature that can sometimes lead to major performance degradation. On January 6, 2026, etraut-openai wrote that the feature had been removed from the latest CLI, and the issue closed (the issue).
The December 2025 scale is a reference for those levels, not a rule for today. Newer versions add levels, so check the picker before you run it.
Effort levels that GitHub user etraut-openai recommended in December 2025:
| Level | Task the recommendation names |
|---|---|
| low | Simple tasks |
| medium | Most coding tasks |
| high | Complex tasks, such as major refactors |
| xhigh | Rare cases, for the most complex tasks |
When a simple task is slow, check three things in order: the level the session is actually running, the speed of your connection, and whether your CLI is current. The first is the quickest to check.
How do I set a custom reasoning effort for Codex?
Set it once in the configuration file, or switch it for one session with /model. A GitHub issue from August 2025 (issue #2715) asked the same thing: a user wanted headless runs at a high level and could not find where to set it. The issue's author found the key in the user configuration file, which lives at ~/.codex/config.toml (Codex config basics), and GitHub user tibo-openai closed the issue, noting that /model can change the level mid-conversation. The line goes beside your model setting:
model_reasoning_effort = "high"Two more ways follow from the same material. In an interactive Codex CLI session, /model switches the model and adjusts reasoning effort, as the models page documents (Codex models). For one run, pass the same setting with --config. In the same issue, GitHub user wbdb showed it alongside the model flag (the issue).
Projects can set their own level too. The config basics page says a trusted project's .codex/config.toml overrides your personal defaults, and that Codex skips project layers in an untrusted project. A repository can therefore change your level without you noticing. Check the project file when results shift from one repository to another. If your team also runs skills across both tools, what carries over when a Codex skill runs in Claude Code covers the parts that move.
When is the effort setting not the cause?
Changing the level does not fix every slow or weak result, and each cause needs its own check. A slow connection adds time at any level, so test your connection before you blame the level. An old CLI can carry a feature that a newer version has removed, so update it and rerun the task. Configuration follows a fixed order: CLI flags and --config come first, then a trusted project's .codex/config.toml, then your user file at ~/.codex/config.toml. If you change only your user file, a project file can still override it, so confirm the level the session actually runs. A vague request does not get clearer with more thinking either. If a task fails at high, read the prompt for the fact it leaves out before you raise the level again. When a Claude Code session keeps losing the same facts, first check whether the fact was saved, whether it reached the current context, and whether it can be recalled. If it was never saved or cannot be recalled, a memory tool is the next step. Agent-memory is a Tigerless Labs project that keeps Markdown memory shared by Claude Code and Codex CLI, and how to give Claude Code memory across sessions covers the approach.
For a task you can check in a minute, a quick run at low followed by your own review can serve better than one long run at high. Reserve high for work where a wrong answer costs more than the extra wait.
How can I see what a higher effort level costs?
Higher effort typically uses more tokens, and the official pages describe that trade-off without giving counts. To see the counts for your own work, you need to see the requests Codex sends. Tigerless Labs' cost-xray captures them for Codex as well as Claude Code. It runs a local capture, attributes tokens and dollars inside each request to the sources that make it up (for Claude Code the split between sources is an estimate), and shows the result in a terminal view. Install is one line on macOS or Linux, it needs no API key and no account, and it captures from the install onward, not past history.
Cost-xray observes. It does not change the effort you run at, and it does not cut your spend by itself. Confirm that the current Codex session appears in cost-xray before comparing token usage and cost across effort levels. Our Cost X-Ray launch post explains the design, and our walkthrough of what Claude Code sends to the API shows the same kind of breakdown for Claude Code.
Conclusion
Codex effort trades time and tokens for more reasoning on each turn, and the result depends on the task and the model. The December 2025 GitHub advice (low for simple tasks, medium for most coding, high for complex work) is a useful historical reference. The current choice comes from the levels your picker offers and from your own test. The best effort level is the lowest one at which your own test task comes out right, and it changes when your tasks or models change.
Your next step takes about ten minutes. Pick one task you already run each week, run it at medium and at high on the same model, and note the time and the result. One run is an observation; keep the level that holds up across several representative tasks. If medium gets it right, keep medium and keep the time you save. For a broader comparison of the two coding tools, our Codex vs Claude Code guide shows where each one fits, and open-source Claude Code plugins worth installing lists tools for the other side of the same workflow.
agent-memory and AutoHarness, which learns skills from Claude Code sessions, are Tigerless Labs projects for keeping what a session learns. Neither one sets the effort level for you.
FAQ
Does a higher effort level always give better answers?
No. OpenAI's Codex docs say higher effort can improve results on complex tasks, but it takes longer and uses more tokens. On simple tasks a lower level may be enough, so test it on your own work before you choose a higher one.
Where do I change the Codex effort level?
Use /model in an interactive Codex CLI session to change it for that session, or pass --config for a single run. To set it for every session, put model_reasoning_effort in your configuration file at ~/.codex/config.toml. A trusted project can override that with its own .codex/config.toml.
Do all models offer the same effort levels?
No. The levels depend on the model and the client you use, and the picker in your app shows the levels you can choose.
Does changing the effort level change the model?
Changing the effort level keeps the model you selected. Changing the model is a separate choice, made with /model or -m. The setting changes how much the model reasons before it answers.
Does cost-xray change how hard Codex thinks?
No. Cost-xray observes the requests and reports the tokens and cost inside each one. It does not change the effort you run at, and it does not cut spend by itself.
Sources
- OpenAI API docs: Reasoning models
- Codex models (OpenAI)
- Codex config basics (OpenAI)
- openai/codex issue #8545: Extremely low performance on simple tasks
- openai/codex issue #2715: How to set custom reasoning effort
- OpenAI developer forum: How to decide a level of reasoning?
- OpenAI developer forum: Can Codex low match medium on easy tasks?
- Tigerless Labs cost-xray
- Tigerless Labs agent-memory
- Tigerless Labs AutoHarness
