What Is ChatGPT Codex?

What Is ChatGPT Codex?

Codex is the name OpenAI has now used for two quite different products, which is the main reason people arrive at this question confused. The original Codex was a code-completion model that powered early autocomplete tools and was retired years ago. The Codex people mean in 2026 is something else entirely: an agent that takes a task, works on a codebase on its own, and comes back with a change you can review.

The distinction matters because it changes what the tool is for. An autocomplete model helps you type faster. An agent does a unit of work while you do something else. Those sit in completely different places in a working day.

This guide covers what Codex actually is now, where it runs, how access is structured across OpenAI's plans, and the specific situations where it is or is not the right tool.

What Is ChatGPT Codex? The Direct Answer

Codex is OpenAI's software-engineering agent. You give it a task in plain language — fix this bug, add this endpoint, migrate this module — and it reads the relevant code, makes changes, runs commands, and presents a result for you to review.

It is agentic rather than conversational. The interaction is not a back-and-forth about code; it is a delegation. You hand over a task and get back a diff, a branch, or a pull request depending on where you are running it.

It launched in April 2025 as an open-source command-line tool and has expanded steadily since: a cloud version in May 2025, a dedicated desktop app in February 2026, a Windows version with PowerShell support in March 2026, and a security-focused variant the same month. In July 2026 OpenAI merged the standalone Codex app into the ChatGPT desktop app, which is why the product now shows up under the ChatGPT umbrella rather than as a separate download.

Trying to work out which AI tools actually earn their place in your workflow? Get a free SEO audit and start with the work that moves your numbers.

How Codex Actually Works

Codex is less a single application than one agent exposed through several surfaces. Understanding which surface you are using explains most of the differences people run into.

The CLI

The original form, and still the one many developers prefer. It is open source under Apache 2.0, runs locally in your terminal, and operates on the repository you are sitting in. Because it runs on your machine, it sees your real environment — your dependencies, your config, your local state.

The Cloud

Cloud tasks run in an isolated environment on OpenAI's infrastructure instead of your laptop. You delegate a task, close the lid, and come back to a finished result. The trade is that the cloud environment has to be configured to resemble yours, which is setup work you only do once but do have to do.

The Desktop App

Introduced in February 2026 to manage several agents at once, and folded into the ChatGPT desktop app in July 2026. This is the surface for running multiple tasks in parallel and watching them progress, rather than babysitting one.

IDE and Platform Integrations

There is a Visual Studio Code extension, plus JetBrains and Xcode support. Codex also reaches into GitHub through Agent HQ and into Figma through its MCP server, which lets a design file act as an input to a coding task.

The Models Underneath

Codex runs on models tuned specifically for software engineering rather than general chat. GPT-5.3-Codex arrived in February 2026, a lower-latency variant followed about a week later, and GPT-5.4 landed for Codex in March 2026. The naming moves quickly; the practical point is that the Codex models are a separate line from the general-purpose ones and are optimised for longer, tool-using work.

How to Start Using Codex

The fastest honest path from zero to a useful result is shorter than most people expect, but the order matters.

Step 1: Pick Your Surface First

Decide between local and cloud before you install anything. If you want Codex working on the code in front of you right now, start with the CLI. If you want to hand off tasks and walk away, start with cloud.

Step 2: Start in a Repository You Know Well

The first task should be one where you can instantly tell whether the output is right. A bug you have already diagnosed is ideal, because you are evaluating the tool, not the solution.

Step 3: Write the Task Like a Ticket, Not a Prompt

Agents do better with scope and acceptance criteria than with conversational instructions. State which files are in scope, what done looks like, and what must not change. Vagueness is where agent runs go wrong.

Step 4: Let It Finish Before Intervening

The common failure in the first week is interrupting halfway because the intermediate steps look wrong. Agents often take a detour and recover. Judge the result, not the path.

Step 5: Review the Diff Like a Colleague's Pull Request

This is the part nobody should skip. Read the change, run the tests, check for things that were quietly edited beyond the stated scope. The review is not a formality; it is where the value is realised or lost.

Measuring whether a new tool actually improved your output is its own problem. Get a free SEO audit and get a baseline you can compare against.

Which Plans Include Codex

Access is tiered across OpenAI's subscriptions rather than sold separately, and the shape of the tiering tells you who it is aimed at.

The Entry Tiers

Free and the low-cost Go tier offer limited, mostly local use — enough to try the tool, not enough to work with daily. Treat these as evaluation rather than adoption.

Plus

The standard working tier. Plus unlocks the full set of surfaces, including cloud task delegation, the CLI, the IDE extensions and mobile. For an individual developer this is the normal starting point.

The Pro Tiers

OpenAI added a mid-weight Pro option in April 2026 alongside the existing higher one, with usage multiples rather than different features. You are buying headroom, not capability.

Business and Enterprise

Business is priced per seat and carries the same base limits as Plus, with the option to add credit bundles. Enterprise runs on a shared credit pool with negotiated terms, and notably treats Codex seats as a separate line item from ordinary ChatGPT seats — worth knowing if you are costing a rollout.

How Usage Is Counted

In April 2026 OpenAI moved from per-message counting to token-based credits, measured over rolling multi-hour windows rather than monthly allowances. Heavier models draw credits faster. Because these rates and limits have changed more than once, check the current figures on OpenAI's own pricing page rather than trusting any number you read in an article, including this one.

The API Route

Codex models are also available through the API without a ChatGPT subscription, billed per token with a prompt-caching discount. API access omits the cloud conveniences such as GitHub code review, and newer models typically reach API users later than subscribers.

When Codex Is the Right Tool

The honest answer is that it is strong at a specific shape of work and unremarkable outside it.

Well-Specified, Tedious Changes

Renaming across a codebase, adding a well-understood endpoint, writing tests for existing behaviour, migrating a dependency. Work where the definition of correct is clear and the effort is in the typing.

Parallel Work You Would Otherwise Queue

The ability to run several cloud tasks at once changes the economics of small jobs. Things that were never worth a context switch become worth delegating.

Unfamiliar Codebases

Asking an agent to trace how something works and summarise it is often faster than reading, and the summary comes with file references you can verify.

When It Is Not the Right Tool

Architectural decisions, anything requiring judgement about trade-offs, work in a domain with rules the model cannot know, and anything where a plausible-but-wrong answer is expensive. Agents are confident in all of these cases and should not be.

Not sure where automation actually pays off in your own workflow? Get a free SEO audit and find the work worth automating.

Limitations and Things to Watch

The limitations are not secrets, and knowing them up front saves a bad first month.

Code quality is the live debate. Security researchers have raised real concerns during 2026 about the quality of AI-generated patches, including fixes that look right and resolve nothing. Review is not optional.

Cloud environments drift from local ones. A task that passes in the cloud sandbox can fail on your machine because the environments differ, and reconciling them is recurring maintenance rather than one-time setup.

Usage limits bite harder than the headline suggests, because they are measured in rolling windows. A heavy afternoon can exhaust a window that a steady week would not.

Model naming changes fast, and so do the limits attached to each one. Anything you write down about tiers today has a short shelf life.

And the product boundary keeps moving. OpenAI has stated an intention to merge its desktop clients into a single application, so where Codex lives is still in flux.

Codex vs the Alternatives

The agentic coding market consolidated quickly, and the differences now are more about workflow than raw capability.

Codex vs Claude Code

The closest comparison, and the competition is direct enough that Reuters reported in April 2026 that OpenAI had redirected resources toward Codex in response. Both are terminal-first agents with strong multi-file reasoning. The practical difference tends to be which model you prefer working with and which subscription you already pay for.

Codex vs GitHub Copilot

Different categories despite the overlap. Copilot grew from inline completion outward; Codex started as an agent. If you mostly want suggestions while typing, Copilot's model fits better. If you want to hand off whole tasks, Codex does.

Codex vs IDE-Native Agents

Editor-integrated tools keep you in one window and make the review loop tighter. Codex trades that for running outside the editor entirely, including in the cloud while you do something else.

Codex vs Doing It Yourself

Worth stating plainly: for a task you could finish in ten minutes and fully understand, writing the ticket and reviewing the diff can cost more than the work. Delegation has overhead, and it only pays above a certain task size.

Final Thoughts

ChatGPT Codex is OpenAI's coding agent — a CLI, cloud service, desktop surface and set of IDE integrations sharing one engineering-tuned model line, included at varying levels across ChatGPT's paid tiers and available separately through the API.

The useful mental model is delegation rather than assistance. It is at its best on well-scoped work you could do yourself but would rather not, and at its worst on anything requiring judgement you have not supplied. The review step is where that distinction gets enforced, and skipping it is how teams end up with confident, plausible, wrong code in production.

If you are weighing it against the other serious option in this space, our comparison of Claude Code and GitHub Copilot covers how the agentic and completion-first approaches differ in daily use.

Try it on a task where you already know the right answer. That is the only evaluation that tells you anything, and it takes an afternoon rather than a trial period.

Want the same clarity about where your site stands? Get a free SEO audit and see what is actually worth fixing first.