Codex

Codex is a product name from OpenAI, the company behind ChatGPT. Behind it are computer programs that generate program code from a description in everyday language. Program code is the set of instructions a computer works from, written in a fixed language such as Python. So you type “sort this list of names alphabetically,” and Codex writes the matching lines. The name has existed since 2021, when it referred to a single model behind the programming assistant GitHub Copilot. Today, Codex refers more to a tool that handles entire programming tasks independently.

Why programming became a testing ground for AI

Programming is a particularly attractive application area for AI companies. The reason is simple: with code, you can automatically check whether the result is correct. You run the program and see whether it produces the right output. With an essay or a piece of advice, that is much harder to measure.

On top of that comes the economic appeal. Software developers earn well, and companies pay accordingly for their working time. A tool that takes over part of this work therefore has a clearly calculable value. This is exactly why Codex and competing products like Claude Code or Cursor appear so often in business news.

A common misconception is that such tools replace programmers. So far, something different is happening: the role is shifting from typing to reviewing. Anyone who cannot read code won’t notice when Codex produces something wrong or insecure. Expertise is thus not becoming obsolete, but is being applied differently.

From text model to agent with a terminal

The foundation is a language model. This is a program that has learned to continue texts by being trained on vast amounts of text. For Codex, a large part of this training material consisted of publicly available program code. This allowed the model to learn what typical solutions to typical tasks look like. It does not understand code the way a human does; it continues patterns with great accuracy.

The second step turns this into an agent, that is, a system that carries out actions itself. Codex is given access to an isolated working environment containing a project’s files. There it can read files, modify them, run tests, and read the resulting error messages. If a test fails, it attempts a fix and tries again. This loop of changing and checking repeats until the task appears to be complete.

In the end, Codex does not deliver a raw piece of text but a proposed change to the project. A human then reviews line by line what was changed and either accepts or rejects it. The comparison to an intern works well here: it works independently, but someone reviews the result before it goes into operation.

Codex in the editor, in the terminal, and in the cloud

Most often, one encounters such systems directly while programming. In the text editor, a faint suggestion for the next lines appears while typing, which you accept with the Tab key. Many know this feature as GitHub Copilot, whose first version was built on Codex. It is now used daily by millions of developers.

Besides that, Codex also exists as a command-line tool and as an online service. You describe a task such as “fix this bug in the login,” and the system works on it for several minutes. For students, access via ChatGPT is the most obvious route: code can be generated and explained there as well. Anyone who needs a formula in a spreadsheet or a small script can get further this way without prior knowledge.

In the news, Codex often serves as an example in two larger debates. The first concerns copyright, because the training used other people’s code. The second concerns security, since an agent with write access to a project can also cause damage. Neither question has been conclusively resolved to this day.