What is Codex in ChatGPT?
Codex is the software-development half of ChatGPT — a separate experience alongside Chat and Work, dedicated to writing and debugging code, running tests and commands, reviewing changes and working with repositories. It lives in the ChatGPT desktop app and is not selectable on web or mobile. Its history is kept separate from your ChatGPT chat history, and its usage is metered in credits rather than messages.
- What it is
- Coding agent code, tests, repos, terminals
- Where
- Desktop app macOS and Windows
- On web / mobile
- No remote access from mobile only
- History
- Separate not in ChatGPT chat history
- Metering
- Credits shared pool with Work
- Extra usage
- Purchasable credits beyond plan limits
- Our coverage
- Full section /agents/codex/
- Verified
- 2026-08-21 OpenAI help centre
It is a different product sharing a login with the one you know.
OpenAI's current documentation puts Codex alongside Chat and Work as one of three experiences, and describes its job precisely: "Write or debug code, run tests and commands, review changes, or work with a repository." Unlike the other two, it is a separate view in the ChatGPT desktop app, reached from the top-left menu, and its workflows and history stay separate from ChatGPT's.
The platform constraint is the thing most people discover the hard way: Codex is not selectable on web or mobile. You can reach supported desktop Codex chats from the Remote tab in the ChatGPT mobile app — approving actions and checking progress from a phone — but those chats never become part of your web or mobile chat history. If you have been looking for Codex on chatgpt.com, it is not hiding; it is not there.
It works where code lives: local folders, repositories, terminals and developer tools, with inline editing within diffs and pull-request review in a side panel. Recent additions include a plugin directory for packaged workflows, computer use, and remote control for Windows hosts. Voice works with Codex on desktop, drawing on the same tools and permissions the experience already has.
Billing is where Codex stops behaving like ChatGPT. Work and Codex share an agentic usage and credit pool, and additional credits can be bought beyond plan limits. That is a genuinely different model from the unlimited-chat-plus-metered-tools structure of the rest of the product — and OpenAI does not publish a dollar price for a credit anywhere, which is why we took its rate card apart separately.
What we know that OpenAI doesn't print.
Codex meters usage in credits. OpenAI publishes what a plan includes and what more credits cost to buy, but not what a credit is worth in dollars of model usage — which makes comparing Codex against any per-token alternative impossible on the vendor's own numbers.
So we did the arithmetic ourselves: divide the published rate card by the per-token prices in our pricing snapshot, and show every line of the working. That is on our Codex pricing page, along with the access routes and what each one actually includes. It is the single most useful thing on this site about Codex, and it is not something you will find in the help centre.
This page answers what is it. The section that answers should you use it, what does it cost, and what does it fail at is /agents/codex/, which covers Codex the way we cover any tool we take seriously: every route in priced from the vendor's own page and stamped with a date, and the honest cases where something else is the better choice on the alternatives page.
One thing we will say plainly here, because it is the kind of claim that is usually fudged: we have not run Codex ourselves. Doing so would require signing into somebody else's ChatGPT account, which we do not do. Where our Codex pages rest on documentation rather than our own runs, they say so on the page.
The comparison people actually want is against Anthropic's Claude Code, and we have it written out at Codex vs Claude Code — plus Codex vs Cursor for the editor-shaped alternative.
The structural differences that matter before any benchmark: Codex lives in a desktop app and bills in credits from a pool shared with Work; Claude Code is a terminal tool covered by a Claude subscription or API credits; Cursor is an editor. Those shapes decide more about whether a tool fits your day than any leaderboard does, which is why our ranking is built from our own comparison pages rather than from a score we invented.
Coding agents we have taken apart.
Each of these has a full section on this site — every access route priced from the vendor's own page, dated, with the cases against it stated. Prices from our verified sources, checked 2026-07-16.
Frequently asked.
Quick follow-ups people search after this question.