Codex

The choice comes down to two profiles: see which one is yours.

Who it is for. Codex is built for a flow of work: many tasks launched at once and almost for free. Hand it a dozen small fixes in the morning, clear out stale changes, review someone else's work — the agent handles several tasks at once, each in its own copy of the project. Sign in with a regular ChatGPT account, even a free one. You can assign tasks from the app, browser, and phone.

Don't write code? A terminal is not required: assign a task in the ChatGPT app or browser, just like in chat. At OpenAI, they themselves call the agent a superpower for people without development experience

Who it is not for. There are no suggestions as you type here - for that, go to Copilot. OpenAI models only, you cannot plug in your own local ones. Design and good-looking screens are taken to Claude Code: AI's weakness in design even the head of Codex admits it.

Just getting started? Haven't written code or built projects yet - start with the page“What is vibecoding”, then come back here for a tool.

If it suits you, install it using section 3. If you're choosing between it and Claude Code, see the decision tree in 7.

Watches the sources. Several times a day, the agent rereads 15–33 pages where AI news emerges and compares them with yesterday's. On August 17, it went through 15 sources and picked nothing — it was all noise. This refusal is work too: junk does not make it into the feed.

Writes entire news stories. The agent put together a story about Lovable starting to deliver apps in ChatGPT itself: a verdict, what changed for the reader, and sources.

Judges every story. A judge stands before the writer, and that is Codex too: it scores every story, separates urgent from ordinary, and decides what is not worthy of the feed.

Draws covers at night. The agent generated four clay-style course covers overnight, all four in a consistent style on the first try. The pipeline orders news covers from it on its own, without a person.

Clicks through panels like a person. When the site's hosting needed a different plan, the agent opened a browser, entered the control panel, and switched settings with buttons. It checked afterward: the configuration was applied.

Fixes and checks code. Dependency updates for security, fixing a broken chart on the ratings page, a new screen in the internal dashboard — the agent handles routine engineering work on its own.

What it looks like. The cycle is simple: you assign a task in words from the ChatGPT app, terminal, or phone — the agent edits files in a sandbox with networking disabled, shows every change line by line, and takes it to a finished pull request. Several such tasks can run at once: in the cloud, the agent makes parallel attempts and shows the best one.

How much freedom to give it. By default, the agent asks for confirmation on risky steps; you choose the boundaries from there — up to a complete “do not ask.” The agent knows the line too: in a case shown by Greg Brockman, Codex assembled and launched an ad campaign, but stopped at the payment button and asked for permission.

Then on its own. Higher up the axis is goal mode: put a goal in front of the agent, not a list of steps. Plus computer use: Codex clicks through your apps on Mac and Windows — writes throwaway scripts instead of live clicks, which is faster. Using this setup, GPT-5.6 Sol in Codex in five hours completed the daily challenge itself in Slay the Spire 2. The top level is the cloud: the pipeline works while your laptop is closed, and you can manage it from your phone. How to hand work back to the agent for another pass — loops.

Dots, the persistent assistant. Since Sep 29, 2026, you can assign a dot, an assistant running on GPT-6 Astra inside ChatGPT, to specific tasks. It works across conversations, keeps context, and can create Work and Codex tasks, continue existing local Codex tasks and check the results of what it delegated. You set the responsibility, for example triaging new bug reports and preparing fixes for review; the assistant coordinates the work.

A dot has its own cloud computer, so cloud work doesn't need your laptop to be on. For code in a specific repository, first create a Codex cloud environment with that repository and its setup. To work with local files, connect your computer separately in the dot's profile: under Computers, find Your computer and click Allow access. Keep the computer online with the ChatGPT app open. Connecting to Codex or Work Sync doesn't grant this access by itself.

Create an assistant in Dots in ChatGPT from a computer. Access is rolling out gradually: Pro $100, $200 and $500 for users over 18 outside the EEA, the UK and Switzerland; Business Premium and Enterprise worldwide, where in Enterprise an admin turns it on. The first dot is included in Pro and Business Premium at no extra cost, but deep work has a quota. Chatting with it doesn't use your ChatGPT limit; Codex and Work tasks use their usual limits. Free access to Codex doesn't mean access to Dots.

Review child tasks in the dot profile's Activity and accept code based on the diff and tests. Pause stops only the assistant's current main task: stop child tasks separately in Activity, and cancel schedules in Scheduled. A finished run doesn't by itself prove you got the result you needed.

Dots: announced Sep 29, 2026, checked Sep 30, 2026 against the task guide, access terms and stopping work. We haven't tested Dots features in this snapshot ourselves.

Where the code goes. On personal plans - Free, Plus, Pro - OpenAI may train models on your content by default: disable training with the toggle in Settings → Data Controls. Business plans and the API do not train by default, while Enterprise adds zero data retention. Turned off the main toggle? Check Codex's own settings too: third-party analyses have found a separate control there.

What the agent can break. It edits files and runs commands - but by default in a sandbox with networking disabled: Seatbelt on macOS, bwrap on Linux, and a system sandbox or WSL2 on Windows. Save "no sandbox" mode for extreme cases: in July, it was in this mode that the agent deleted a user's files while overriding the home directory - the bug was acknowledged and fixed.

The agent checks its own code. The CLI is open source - you can read the code to see exactly what the agent does. And since August, Codex has reviewed your pull requests for vulnerabilities too: Security Review in preview, it comments only on P0 and P1 risks - in an independent test, it caught 16 of 23 known vulnerabilities.