OpenAI Codex is an AI coding agent: you describe a software task in plain language, and it reads the relevant files in your project, writes the code, runs it, and shows you the result as a diff to approve or reject. It is not an autocomplete tool that finishes your sentence — it is meant to carry out a whole task and come back with working code.
What Codex actually is
Codex is built by OpenAI, the company behind ChatGPT. The name is recycled from an older 2021 OpenAI model that once powered early versions of GitHub Copilot, but today’s Codex is a different product: a coding agent OpenAI relaunched in 2025, first as a command-line tool (Codex CLI) and then as a cloud-based agent (Codex Cloud). You can run it from a terminal, inside an IDE such as VS Code or JetBrains, through a web browser at chatgpt.com, or from the ChatGPT desktop app.
Every ChatGPT plan includes some Codex access, since it is no longer a separate paid product — it sits alongside regular chat as one of the things ChatGPT can do.
How it works
When you give Codex a task — “fix this failing test,” “add a login page,” “upgrade this dependency” — it opens an isolated environment preloaded with your repository (locally through the CLI, or in an OpenAI-hosted cloud sandbox for longer jobs). Inside that environment it can read files, edit code, run your test suite, and call other command-line tools, the same way a developer would work through a ticket. Most tasks finish in one to thirty minutes. When it’s done, Codex returns a diff, along with the commands it ran and their output, so a human can review exactly what changed before merging it.
Because the agent can act — not just suggest — it typically runs with limited permissions and inside a sandbox, and any command that reaches outside that sandbox (like pushing to a remote repository) normally requires explicit approval.
How it differs from a code-completion tool
Tools like GitHub Copilot started out as inline autocomplete: they suggest the next few lines while a developer types. Codex is built the other way around — you hand it a self-contained task, it works through it independently in the background, and you review the outcome once. Many coding assistants, including Copilot, have since added their own agent modes that work similarly, so the line between the two styles has blurred; Codex was OpenAI’s entry into that agent category.
What it costs
Codex usage is bundled into ChatGPT’s regular pricing rather than sold as a stand-alone subscription. As of July 2026, per OpenAI’s pricing page: a free tier gives limited access; Go is $8/month; Plus is $20/month; Pro starts at $100/month for 5x the usage limits (or $200/month for 20x); Business runs $20 per user/month on annual billing for teams of two or more; and Enterprise/Edu pricing is custom. Usage is metered in credits tied to how many tokens a task consumes, on a rolling five-hour window, and developers who prefer to pay per use can connect an OpenAI API key instead of a ChatGPT plan. Because these figures change, check OpenAI’s current pricing page before budgeting around them.
Getting started
To try Codex, sign in to the ChatGPT desktop app or chatgpt.com with a ChatGPT account, connect a GitHub repository, and describe a task. The CLI version installs with a single terminal command for developers who prefer to stay outside a browser.
In the news
On July 9, 2026, OpenAI folded Codex into a single ChatGPT desktop app alongside a new agentic mode called ChatGPT Work, rather than keeping it as a separate application — read our report on the ChatGPT Work launch for what changed.
FAQ
Is Codex the same thing as the 2021 Codex model that powered early GitHub Copilot? No. It shares a name and some heritage, but the current Codex is a separate agent product OpenAI rebuilt from 2025 onward.
Do I need to know how to code to use Codex? You need enough understanding to review the diff it produces and judge whether the change is correct — Codex writes the code, but a human is still expected to approve it.
Can Codex work without an internet connection? No. Even the CLI version needs to reach OpenAI’s models to generate and reason about code, though it can execute and test the resulting code locally.
Does Codex replace developers? It automates specific, well-defined tasks under human review; it does not remove the need for someone to decide what to build or to check the result.