Chat GPT Codex Review (2026): Access, Limits and Cost

What chat gpt codex is in 2026: one coding agent that runs in the ChatGPT app, a desktop app, the CLI, your IDE and the cloud, on the plan you already pay for.

Try Begin free → Official site

What chat gpt codex is now

The name has moved twice, which is why the search results are confusing. Codex started in April 2025 as a coding agent for developers. In February 2026 OpenAI shipped the Codex app for macOS, described as a command center for agents rather than a chat window: you run several agents in parallel, each on its own task, and supervise work that can span hours. Windows support followed on 4 March 2026. The same agent is also reachable from inside ChatGPT, from the CLI, from an IDE extension and from the cloud, and OpenAI states that your rate limits apply across all of those surfaces rather than per tool. By July 2026 the company said more than five million people use Codex every week, and that more than a million of them use it for work that is not software at all.

Which plan gets you in

When the app launched, OpenAI said it was including Codex with ChatGPT Free and Go for a limited time, and doubling rate limits on Plus, Pro, Business, Enterprise and Edu. Those higher limits follow the account, so the same subscription covers the app, the CLI, the IDE plugin and cloud runs. Two things follow from the wording. First, free access is a promotion with no stated end date, so a workflow built on it can become a paid workflow without warning. Second, the useful comparison between tiers is not features but headroom, because every surface draws from the same allowance. A per-plan breakdown lives on the Codex pricing page, and the developer documentation sits on the OpenAI developers site rather than in the consumer help centre.

Skills, automations and parallel agents

Three capabilities distinguish the app from a chat session. Skills extend the agent past code generation, so it can be pointed at work that has a repeatable shape. Automations take repetitive jobs off your hands on a schedule rather than waiting for a prompt. Parallel agents let you start several tasks at once and check them like a build queue, which changes the unit of work from a message to a job. There is also a configurable personality setting, which sounds cosmetic and is not: an agent that asks fewer questions behaves very differently on a long task. OpenAI describes the security posture as secure by default and configurable by design, which in practice means sandboxing and approvals you can loosen deliberately rather than by accident.

Who is running it in production

The Codex page lists Heidi, NVIDIA, Cisco and Shopify among the teams using it, which tells you the target is a working engineering organisation with review processes, not a weekend project. That framing matters when you evaluate it. An agent that opens pull requests is only as useful as the review culture around it, and the pitch is explicitly about routine pull requests scaling into larger engineering work. If you have no repository, no CI and nobody to review a diff, most of what Codex is good at has nowhere to land. The July 2026 announcement of ChatGPT Work, which has Codex technology built in, points the other way: the same engine aimed at documents, sheets and slides for people who never open a terminal.

When a coding agent is the wrong tool

There is a category of request that looks like programming and is really assembly: a landing page, a small marketing site, a copy of a layout you already like, a simple mobile app to hand to a client. Codex will do it, and then you still own the repo, the build step and the hosting decision. Begin.sh takes the other route. You give it a prompt, or a URL to clone, and it returns a working static site or an Expo app as a zip you download. No hosting, no backend, no auth, which is a real limitation if you need accounts or a database and exactly the point if you do not. Use the agent for the codebase you maintain, and the generator for the artefact you just want to exist.

What you actually get

Codex against a one-shot generator

Getting a useful answer out of it in an hour

1. Check your plan first
   Open the Codex pricing page and confirm what your current ChatGPT tier includes. Free and Go inclusion was announced as limited-time, so do not assume today's access is permanent.
2. Pick one surface
   Start in the desktop app if you want to watch agents work, or the CLI if you already live in a terminal. Limits are shared, so switching later costs you nothing.
3. Give it a scoped task
   A failing test, a small refactor, a dependency bump. Long autonomous runs are the selling point, but a bounded first task tells you more about review quality.
4. Read every diff
   Treat output as a pull request from a fast new hire. The teams named on the product page run it inside real review processes, and that is the part that makes it safe.

FAQ