Codex
Codex is an artificial intelligence agent from OpenAI that writes, fixes and reviews software code. It works in the ChatGPT app, in the cloud, in code editors and in the terminal. Tasks are described in plain language, and Codex makes the changes in the repository itself and proposes them for review. The name Codex previously belonged to a 2021 language model that powered the first version of GitHub Copilot.
What is Codex?
Codex is a coding agent, i.e. an AI tool that carries out tasks in code on its own rather than only suggesting individual lines. OpenAI describes it as an agent for real engineering work: from finding bugs, through refactoring, to new features and code reviews. Codex uses a ChatGPT account, and the same work is available in the app, the editor and the terminal.
| Feature | Description |
|---|---|
| Creator | OpenAI |
| Type | AI agent for writing and reviewing code |
| How it works | tasks described in natural language, repository changes for approval |
| Where it runs | ChatGPT, Codex cloud, code editor extension, Codex CLI in the terminal |
| Agent instructions | AGENTS.md files in the repository |
| Access | included in paid ChatGPT plans |
| Related terms | GitHub Copilot, ChatGPT, language model |
Codex differs from an ordinary AI chat in that it has access to the project's code and can run it. The agent reads files, makes changes, runs tests and shows the result for review. A human decides whether the changes go into the project.
How has Codex changed since 2021?
The name Codex first referred to a language model, and since 2025 it has referred to a coding agent. Both products come from OpenAI, but they work differently.
| Date | Event |
|---|---|
| 3 June 2025 | Codex becomes available to ChatGPT Plus users |
| 16 May 2025 | OpenAI introduces Codex as a cloud-based agent powered by the codex-1 model |
| 10 August 2021 | OpenAI releases the Codex model in the API as a private beta; the model already powers GitHub Copilot, launched with GitHub a month earlier |
The 2021 model was a descendant of GPT-3. It was trained on text and billions of lines of publicly available code, including from GitHub repositories. It performed best in Python, but it also knew over a dozen other languages, including JavaScript, PHP and TypeScript.
The 2025 agent ran on the codex-1 model, a version of OpenAI o3 tuned for software engineering. It ran each task in a separate cloud environment preloaded with the user's repository.

How does Codex work?
Codex receives a task described in words, e.g. “fix the bug in the contact form”, and carries it out in the project's code. It can write new features, answer questions about the code, fix bugs and prepare changes for review as a pull request. According to OpenAI, it can also run multiple agents in parallel and check its work with computer and browser tools.
| How to use it | Where it runs |
|---|---|
| Codex in ChatGPT | in the ChatGPT app |
| Codex cloud | in OpenAI's cloud, tasks keep running after you close your laptop |
| Editor extension (IDE) | in the developer's code editor |
| Codex CLI | locally on your computer, in the terminal |
| Code review | automatic reviews of pull requests in the repository |
Codex CLI is the version that runs locally, and its source code is public on GitHub. On Mac and Linux it is installed with the command npm install -g @openai/codex. The agent's behavior in a given project is set in AGENTS.md files, e.g. how to run tests and which coding rules to follow.
Codex also connects to other tools through plugins. OpenAI lists GitHub, Slack, Gmail, Google Drive, Figma and Notion, among others. Companies using ChatGPT Business Hub have Codex in the same workspace, with admin controls.
What is Codex useful for in marketing and SEO?
Codex is useful wherever marketing requires changes to the code of a website or online store. Instead of waiting for a developer, the team can ask the agent to prepare a fix, and the developer only reviews it.
- implementing Schema.org structured data in page templates;
- performance fixes flagged by PageSpeed Insights or Google Lighthouse;
- changes to CMS themes or plugins, e.g. new fields, redirects or page templates;
- data analysis scripts, e.g. combining exports from advertising and analytics tools.
The agent does not take over responsibility for the code. Changes prepared by Codex must be reviewed and tested before publishing, just like the work of any developer.
How much does Codex cost?
Codex has no separate subscription, as it is part of ChatGPT plans. As of November 2026, according to OpenAI's website:
| Plan | Price | Codex in the plan |
|---|---|---|
| Plus | USD 20 per month | a limit for focused coding sessions each week |
| Pro | USD 100 per month | higher limits for full days of work across many projects |
| Business | USD 20 per user per month | shared workspace with admin controls, with 2+ seats and annual billing |
Let's talk growth
Send a brief or drop your contact. We'll reply within 24h.
Get a quote