Log inGet started
← Alle Playbooks
Playbook· migration

From ChatGPT to Codex, guide in 10 steps

Since the July rebuild, Codex sits inside the ChatGPT app and is included in every plan, the free one too. Ten steps from first task to habit, with the limits that actually bite.

Codex is the tool most ChatGPT users have and never open. It isn't ChatGPT under another name, it executes real tasks instead of just answering. And since July 2026 it sits somewhere else than most guides on the internet will tell you. Here's the current route, in ten steps.

1. Find Codex, after the 9 July rebuild

If you're reading an older guide, it will send you looking for a separate Codex app. That app is gone. On 9 July 2026 OpenAI folded three things into a single desktop app for Mac and Windows, and Codex has lived there ever since as its own area next to Chat and Work. The old standalone ChatGPT app is now called ChatGPT Classic.

Three ways in: the new desktop app, ChatGPT in the browser, or the extension in your editor. If you don't want to install anything, use the browser and look for the Codex area in the left menu.

The second half of the rebuild matters more: Codex is now included in every plan, the free one too. The line "you need at least Plus for that", which appears in a lot of guides, stopped being true in July. On the free plan you get the middle model tier and a smaller allowance, but you get in.

A tip: don't bookmark a sub-page. The earlier version of this playbook recommended exactly that, reasoning that "the UI shifts but the path stays". That was wrong. The rebuild moved paths too, the developer documentation for instance now redirects to a new address. Remember the name of the area, not the URL.

On models: Codex runs on OpenAI's 5.6 generation, which comes in three tiers. Sol is the strongest and explicitly built for coding and multi-step work, Terra is balanced everyday use, Luna the fast cheap tier. If you push tasks through all day, a smaller tier often serves you better, because you aren't standing at the limit after three hours. Which model for what is in Model landscape 2026.

2. What Codex does differently from ChatGPT

ChatGPT answers you. Codex executes. You give it a real task, Codex spins up an isolated container in the cloud, does the work, returns the result. It's like ChatGPT with hands.

A tip: the first task should be boring and concrete. Summarize a PDF, rewrite a long email, sort an Excel list. Anything you'd otherwise hand to an intern.

3. The first run in under five minutes

Inside the Codex area you have two modes side by side: asking, and having it run. The labels have changed over time, for a long while they were "Ask" and "Code". The distinction stays the same: asking behaves like ChatGPT and answers you, running spins up a container and actually does the work. Take a real thing you did by hand this week, paste it in, pick the mode that runs, wait.

A tip: be concrete about output format. "Bullet list" is weaker than "table with three columns: task, owner, due by". The more precise you specify how you want it, the more usable the result.

4. Where Codex lives

You get Codex in the browser, in the desktop app for Mac and Windows, as an extension in your editor, as a command line tool in the terminal, and in the cloud for runs that keep working without you.

One thing trips people up: since July the desktop app is no longer a Codex app of its own, it's the ChatGPT app, in which Codex is one of three areas. Anyone hunting for a separate download is hunting for nothing. The command line tool, on the other hand, lives on as its own thing, that one wasn't folded in.

One thing worth separating cleanly here, because it gets mixed up a lot. There are two different questions: where does the work run? and what leaves your machine?

The cloud variant computes on OpenAI's machines and never touches your computer. The terminal tool and the editor extension run on your side and can genuinely change your files and execute commands. That is the difference in where things run.

The data question is a separate one: in every variant, whatever content the model needs goes to OpenAI. Running locally does not mean processing locally. Anyone equating "I keep the files local" with "nothing leaves my machine" is mistaken, and with confidential material that mistake is expensive.

A tip: for the first week the browser is enough, nothing to install. The desktop app pays off once you run several tasks side by side. One surface on day one is plenty.

5. Use Codex for non-code things

This is the step most skip. The name sounds like "for programmers only", but in practice I use Codex more for Excel cleanup, PDF summaries, research reports with sources, translations with consistency checks, Notion reorganizations, and bug-report sorting than for actual code.

A tip: write down three tasks you do by hand every week and don't want to anymore. Hand them to Codex for two weeks. One never lands on your desk again.

6. Understand the limits before you pay

Codex costs nothing extra, it sits inside the plan. That's the easy half. The half people actually get stuck on is the shape of the limit, and that matters more than any number.

There are two caps at once, not one. A rolling window across a few hours, and a weekly cap on top of it. So you can do everything right on Tuesday, never touch the hourly cap, and still be stuck on Friday because the week is full. Anyone holding only the smaller cap in their head plans wrong.

The second lever is the model tier. The same task costs you a fraction of the allowance on the fast tier compared to the strongest one. For routine work the smaller tier isn't the compromise, it's the right call.

I deliberately don't quote concrete allowances here. They've changed several times this year, most recently in early August, and a number in a playbook is correct exactly until the next adjustment. The current values are at OpenAI itself.

A tip: start free, now that you can. Two weeks of real use, counting how often you actually hit a cap. Upgrade after that, and to the tier your real consumption points at. Upgrading as a precaution is the most expensive way to answer this question.

7. Connect GitHub only if you really have code tasks

Non-code tasks don't need GitHub. The connector exists so Codex can read, review and change repositories, but that's not needed in the first month. Stick with single file uploads, that's enough for 90% of all uses.

A tip: if you want to test code tasks for real, attach a small private test repo first. Have Codex review a tiny change, see how it feels before letting it loose on something important.

8. Solve the memory problem

This is the biggest trap when switching from ChatGPT. ChatGPT builds up a picture of you over time, Codex doesn't. A new thread has no personal long-term memory that grows with you. Project knowledge, on the other hand, you can store, and that's the lever. Two honest solutions. The simple one: put a file called AGENTS.md in your work folder and write your project facts into it. Mind the spelling, Codex looks for exactly that name and reads it on startup by itself. A file called anything else you have to mention by hand every session, and that is precisely what you forget. The complete one: attach external storage via MCP. MCP is a protocol Codex understands natively, so your facts and decisions are simultaneously available in Claude, Codex, and your other tools.

A tip: start this week with the Markdown file. That catches 80% of the pain. Get into MCP only once you really feel the gap.

If you're coming from another tool, check your Codex version's command list before you start retyping: at some point there was an import for existing notes from Cursor and Claude Code. Whether it's in your version, and what it's called, the application tells you more reliably than this playbook can. You need the rest of this step either way, because imported isn't maintained.

9. Accept the cloud execution model

Each time you click "Code", your task runs in an isolated cloud container that's discarded after completion. That's why tasks take one to two minutes, and why you don't have to worry about Codex touching your local machine when you use it in the browser. The result stays, the container doesn't.

A tip: don't stare at the spinner. Start the task, open another tab, switch to your phone, come back when done. Codex is built async, use it that way.

10. Build a small habit, not a giant project

The fastest way to get Codex into your daily life: replace one recurring manual task with it. Not ten, not a whole workflow, just one. Monday morning inbox to action list. Friday weekly report. Wednesday KPI cleanup. Take the smallest recurring thing you really do and hand it to Codex for two weeks.

A tip: when the first habit sits, add the next. Stacking is what turns a tool you opened once into one you rely on.

What's next

If you've done these ten steps, you know Codex well enough to decide if it stays in your setup permanently. The honest weakness is the missing memory between sessions. The honest fix, without you having to buy something new, is the MCP path.

If you want to go further there, see the playbook "Use memory portably" or directly Level 4 of the Academy. There we show how external storage works simultaneously in Codex, Claude, and Cursor. One source of truth, many clients. That's what Codex alone doesn't offer.

For now: open it, throw in a task, two weeks straight. The rest comes by itself.

Sources

Surfaces and plans change faster at OpenAI than any guide can keep up with. The description above reflects 9 August 2026. Where a number would go stale quickly, this playbook deliberately states the mechanism instead.

← From ChatGPT to Claude, guide Your first MCP server in 90 mi→
From ChatGPT to Codex, guide in 10 steps — StudioMeyer Academy