ChatGPT and Codex for programming: what OpenAI offers
OpenAI currently offers essentially two different ways to help with coding, and people regularly mix them up because both are tied to the same company and the same subscription. The first is a regular chat with ChatGPT, where you describe a task, get a snippet of code in the browser window, and paste it into your project yourself. The second is Codex, a separate agent tool that works not in a chat, but directly inside the repository: it reads files, runs commands, and makes edits on its own. These aren't just two names for the same thing — they are different workflows with different levels of autonomy.
For a business deciding whether to hire a developer, tackle the task in-house, or hand it off to a freelancer armed with AI tools, the difference between these two modes directly impacts what you can realistically expect from the result and who needs to review it before it hits the production system. Below — no hype, just facts — is what each of these modes can actually do in practice, how much it costs, and where you need to be careful.
Talking to ChatGPT about code is not the same as an agent in the repository
The classic workflow with ChatGPT looks like this: a person formulates a question, gets a code snippet in the chat window, manually copies it into their editor, runs it, and if there's an error, copies the error text back into the chat — and the cycle repeats. In this mode, ChatGPT doesn't see the entire project structure, doesn't run the code itself, and knows nothing about the other files unless you manually paste them into the conversation. This is convenient for an isolated question or a short script, but it scales poorly to a real project with dozens of files.
Codex works differently: it's a separate product for writing and editing code that works not in a chat window, but directly with the project — via the terminal (Codex CLI), a code editor extension, or a separate cloud environment where you can send a task and get finished changes without having to be involved at every step. Formally, it's still a tool from OpenAI, but in terms of how it's used, it's closer to a developer's assistant than a browser-based chat.
What Codex can do in practice
In command-line mode, Codex receives a task in plain language and then reads the necessary project files, makes edits, and runs commands and tests on its own — basically doing what a person would have to copy and paste manually when working through a chat. Codex CLI is open-source and free to install: you don't pay for the program itself, but for the use of the model powering the agent.
There's also a separate cloud mode: you can send a task to an isolated environment where a copy of the repository is pre-loaded — there, the agent handles the task independently, and the result is returned as a ready-made set of changes for review. The same agent can automatically review pull requests and leave comments on them. Which specific model powers Codex at any given moment is something OpenAI changes more often than an article can reliably track; the principle itself is what matters — the agent works with the project directly, not through the clipboard.
How much it costs and how access works
Codex isn't sold as a separate subscription — access comes with your ChatGPT plan, and the agent usage limit increases with the plan's price. On the free and cheapest paid tiers, access is heavily restricted; the Plus plan costs around $20 a month; Pro is around $100–$200 a month with significantly higher limits; for teams, there's Business at roughly $20–$25 per user; and for large companies, there's Enterprise with custom pricing.
For businesses with uneven agent workloads, there's an alternative to a fixed per-seat price — pay-as-you-go based on actual token usage, without being tied to the number of seats. Keep in mind that in 2026, OpenAI has already revised its pricing model — switching from per-message to per-token billing — so before choosing a plan, it's smarter to check the current pricing page rather than relying on a figure from a six-month-old article.
Who this is already a good fit for, and who should wait
Agent mode makes sense where you already have someone who knows how to work in a terminal or code editor and can formulate a task precisely enough: fix a specific bug, write a test, or do a clear module refactoring. For a short, isolated question or a one-off script, a regular ChatGPT chat is still a simpler and faster tool — not every task requires an agent digging into the repository.
Fair warning: in cloud mode, the agent can edit multiple files at once without step-by-step confirmation for each action — this is what makes it fast, but also what makes it risky. For businesses whose code handles payments or customer data, this means the output must be reviewed by a human before it goes live, rather than being taken at face value just because the agent reported the task as done.
Frequently asked questions
Now that Codex is here, is regular ChatGPT still needed for coding?
It is still needed for other tasks. For a quick, isolated question, explaining someone else's code, or a one-off script, the chat remains faster and simpler than spinning up an agent to work with an entire repository.
Is Codex free to use?
The Codex CLI itself is open source and free to install, but running it counts against your ChatGPT plan's model usage limits or is billed by the token — so it won't be entirely free once you start using it regularly.
If the agent writes and runs the code itself, do you still need to review anything manually?
Yes. The agent completes the task as it understands it, but isn't responsible for the business consequences of a mistake. Before changes reach the live system, especially where payments or customer data are involved, a human capable of evaluating the result must review them.