Short answer: I use Low for quick work, Medium for normal business operations, High for complex workflows, and Extra High for the work I cannot afford to have Codex misunderstand.
A Codex reasoning level is the thinking budget Codex uses before it responds or acts.
The mistake is treating reasoning level like a quality score.
That is not how I use it.
I treat reasoning level like an operating budget. I decide how much thinking room Codex should spend based on the business risk, the number of systems involved, and the cost of being wrong.
OpenAI's Codex Prompting Guide says Medium is a strong all-around setting because it balances intelligence and speed. It also points High and XHigh reasoning at the hardest tasks and long-running autonomy. That matches how I use Codex inside a business.
My simple operating rule
- Low: use it when the task is simple and low-risk.
- Medium: use it for most daily business work.
- High: use it when Codex needs to reason across several moving parts.
- Extra High: use it when Codex is working through a high-risk system, a long workflow, or a decision that could affect revenue, customers, data, or infrastructure.
I do not use the biggest setting just because it sounds better. I use the setting that matches the job.
When should business owners use Low reasoning in Codex?
Low is for quick, contained work.
I use it when I already know what I want and I only need Codex to execute a small task.
- Rewrite this paragraph in my voice.
- Give me five title options.
- Summarize this short note.
- Format this checklist.
- Clean up this email before I send it.
- Rename a button.
- Find one obvious typo.
Low is useful because it keeps the workflow moving. I do not need Codex to deeply investigate a sentence rewrite.
When should business owners use Medium reasoning in Codex?
Medium is my default for normal business operations.
I use Medium when the work needs actual thought, but it does not need a full investigation.
- Draft a blog post from a structured outline.
- Turn a transcript into an SOP.
- Review a landing page and suggest fixes.
- Organize lead notes by service type.
- Update a simple website section.
- Write a client-facing guide from bullet points.
- Review a short automation flow.
- Create a workflow checklist for my team.
Medium is the setting I would leave on most of the day. It gives Codex enough thinking room to be useful without forcing every task into a slower, higher-token run.
When should business owners use High reasoning in Codex?
High is for cross-system work.
I use High when Codex has to inspect context, compare files, reason through tradeoffs, and avoid breaking something.
- Audit why leads are not reaching sales.
- Review a full onboarding workflow.
- Fix a website issue across multiple files.
- Compare a local preview against production.
- Check whether an automation is writing to the right system.
- Map a customer journey from opt-in to purchase to access.
- Prepare a launch checklist where missing one step matters.
High is where Codex becomes more like an operator. I want it to slow down enough to understand the system before acting.
When should business owners use Extra High reasoning in Codex?
Extra High is not my everyday setting.
I save it for the work that needs sustained reasoning, careful context handling, and fewer assumptions.
- Redesign a business workflow that affects customers.
- Audit a broken revenue or fulfillment system.
- Plan a migration across tools.
- Review a live launch path before publishing.
- Diagnose a hard bug that crosses several systems.
- Design an agent operating system or approval workflow.
- Let Codex work in Goal Mode on a large, multi-step objective.
Extra High can be worth the token burn when the cost of being wrong is higher than the cost of thinking longer.
How token burn changes the decision
Token burn is not just the visible answer.
Codex uses tokens for the prompt, files, screenshots, command output, context, drafts, and the reasoning it does before responding. Higher reasoning levels can use more thinking tokens because Codex is spending more effort working through the task.
That is useful. It is also not free.
So I ask one question before I change the setting:
How expensive is a bad answer?
If the answer is "not very," I keep the setting lower.
If the answer is "this could break a workflow, confuse a client, miss a lead, publish the wrong page, or waste a team member's time," I raise the reasoning level.
Reasoning levels for automations
Automations need a different mindset because Codex may run without me watching every step.
For lightweight monitoring, I use Low or Medium.
Example: "Check new form submissions every morning and group them by service type."
For automations that touch customers, revenue, CRM data, or access, I use High.
Example: "Review new paid customers, verify access was granted, and flag anyone missing onboarding."
For automations that can affect external systems or long-running business processes, I use Extra High and keep approval gates in place.
Example: "Audit the full buyer-to-member workflow and identify where purchases fail to become access."
Reasoning levels for Goal Mode
Goal Mode changes the stakes because Codex is not just answering one prompt. It is pursuing an objective.
I use Medium for focused goals.
Example: "Turn this transcript into a blog draft, title options, and a LinkedIn post."
I use High when the goal touches multiple files, tools, or decisions.
Example: "Audit this landing page, check the CTA path, update the copy, and verify the live route."
I use Extra High when the goal is broad, operational, or risky.
Example: "Audit my onboarding workflow, find the bottlenecks, and produce an implementation plan with the exact systems involved."
Reasoning levels for agents
Agents are not magic employees. They are only useful when the business gives them the right context, boundaries, tools, and approval rules.
Reasoning level is part of that operating design.
For a simple helper agent, Medium may be enough. For an agent reviewing business operations, High is usually a better starting point. For an agent that has to investigate across systems, preserve context, and avoid bad assumptions, Extra High may be the right choice.
The question is not "how smart can I make this agent?"
The question is "how much reasoning does this agent need to complete this work safely?"
My practical decision table
| Setting | Best for | I avoid it when |
|---|---|---|
| Low | Tiny edits, short summaries, quick formatting, obvious tasks | The work touches customers, revenue, code, or systems |
| Medium | Normal business work, content operations, simple workflows, scoped changes | The task has many dependencies or a high cost of being wrong |
| High | Audits, debugging, multi-step workflows, launch checks, system reviews | The task is tiny or purely cosmetic |
| Extra High | Goal Mode, complex automations, migrations, deep debugging, high-risk operations | The work is simple and speed matters more than depth |
My business-owner rule
I do not choose reasoning level based on ego.
I choose it based on operational risk.
If I am asking Codex to help me think through something that affects customers, revenue, systems, or public pages, I give it more reasoning room.
If I am asking Codex to polish a sentence, I do not.
That is the real use of Low, Medium, High, and Extra High in Codex.
FAQ
Should business owners leave Codex on Medium?
For most daily work, yes. Medium is usually the best default because it balances depth, speed, and token use.
Is High better than Medium in Codex?
High is not automatically better. It is better for complex, risky, or multi-step work. Medium is better for normal operating tasks.
When should I use Extra High?
Use Extra High when Codex is working on a large goal, a complex automation, a migration, a deep audit, or a business process where a bad assumption could cost real time or money.
Does higher reasoning burn more tokens?
It can. Higher reasoning levels can spend more tokens on internal thinking, especially when Codex has to reason through a long or complicated task.
Sources
- OpenAI: Codex Prompting Guide
- OpenAI: Reasoning models guide