Codex

advancedspreading

Codex is OpenAI's coding agent: it independently writes, fixes, and tests code based on text requests, then presents the results for approval.

In plain terms

Codex is like a diligent programming intern to whom you explain, in Hungarian or English, what you want. It reads the existing program, writes the missing part, tests it, and shows you what it has come up with.

It differs from a chatbot in that it not only gives advice but also performs the work. It opens files, runs commands, searches for errors, and can work on larger tasks alone for hours.

The name appeared in the news twice. In 2021, a code-writing language model was called this; since 2025, it has been OpenAI's full coding agent product family. Today, it almost always refers to the latter.

How it works

Describe the task in plain language, for example: 'fix the error that occurs during login with long passwords.'
The agent reviews the project files to understand what is where and how they are connected.
It creates a plan, then modifies the code step by step; it does this on your own machine or in an isolated cloud environment (sandbox).
It runs the tests, and if something breaks, it tries again on its own.
At the end, it summarizes what it changed, and you decide whether to accept it, have it redone, or discard it.

An everyday example

A small online store owner asks their developer to add a new shipping method to the order form. The developer assigns the task to Codex, works on something else, and then reviews and approves the completed modification twenty minutes later.

Who is affected

developerYour work increasingly consists of task delegation and oversight, as the agent handles routine coding. → Code

entrepreneur, executiveSoftware can be developed faster with a smaller team, but oversight and responsibility remain with you.

jobseeker, new graduateAgents now perform some of the tasks junior programmers used to do, so understanding and reviewing code matters more than typing it. → Work & HR

parent, student, teacherIt can be a useful aid while learning programming, but if you have it write everything, you do not learn the thinking process itself. → Education & Family

A common misconception

MythWith Codex, anyone can build finished, reliable software without programming knowledge.

In realityIt delivers working code quickly, but only someone who understands what they are seeing notices bugs, security vulnerabilities, and poor design decisions. The person who accepts the result is responsible for it.

Limits and risks

It confidently writes incorrect or unnecessarily complex code, so every modification must be reviewed before deployment.
It accesses project files and executes commands, so pay attention to what permissions you grant it and what passwords or keys are within its reach.
In large, old, poorly documented systems, it can easily misunderstand relationships.
The command-line program is open-source, but using it typically requires a paid ChatGPT subscription or pay-per-use API access, and both have limits. Your code also passes through an external company's servers.

Similar, but not the same

ChatGPT – ChatGPT is a general conversational partner, while Codex specifically executes programming tasks independently within a project's code. The two are connected: Codex is accessible via a ChatGPT account and subscription.
GitHub Copilot – Copilot is the coding assistant of GitHub (owned by Microsoft). It was originally built on the 2021 Codex model, but it is a separate product and now works with models from several companies.
Claude Code – A similar coding agent, but built by Anthropic and powered by Claude models.

Related concepts

Broader

coding agent, AI-agent, generative AI

Narrower

Codex CLI, GPT-5-Codex

On the edge

Claude Code, GitHub Copilot, Cursor, vibe coding, sandbox

A brief history

OpenAI introduced the Codex coding model in 2021, which powered the first version of GitHub Copilot; this old model was discontinued in 2023. The name was revived in 2025: the command-line Codex CLI appeared in April, the cloud-working Codex agent in May, and the specifically tuned GPT-5-Codex model in autumn.

Where things stand

At the end of 2026, Codex is one of the best-known coding agents, a daily tool for development teams in the command line, in editors and in the cloud. Outside the developer community few people use it yet, and companies are now setting up their rules for oversight.

Status: 2026 November

In the news

Appears in 73 Agent Age stories; the curve shows weekly occurrence in the news magazine.

Google Cloud adds OpenAI models and Codex agents to Model GardenNov 6, 2026 · Code
Postman added itself to Claude Code and Codex without asking permissionNov 5, 2026 · Code
OpenAI and JPMorgan Chase to deploy Codex and cybersecurity modelsNov 5, 2026 · Code
JPMorgan Chase partners with OpenAI on AI for cybersecurity and engineeringNov 5, 2026 · Money
OpenAI to acquire Pierre Computer to strengthen CodexNov 4, 2026 · Code
OpenAI builds new Git backend Pando after Codex agents drive commit surgeNov 4, 2026 · Code
OpenAI may stop supplying models to Cursor from 12 NovemberNov 2, 2026 · Code
OpenAI to pool ChatGPT and Codex usage allowance from DecemberNov 2, 2026 · Code

Also known as

Codex · OpenAI Codex · Codex CLI · Codex agent · Codex agent · GPT-5-Codex · ChatGPT Codex · Codex coding agent

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.