What is GPT-5-Codex, and how does it change OpenAI's coding agent?

Yesterday, Monday, September 15, 2025, OpenAI released GPT-5-Codex, an edition of GPT-5 tuned specifically for software work inside Codex, its coding agent. A coding agent is an AI that does more than suggest a line of code: it reads a project, edits files, runs commands and tests, and keeps going until a task is finished.

GPT-5-Codex is now the default for tasks you send to Codex in the cloud and for code review. Developers working on their own machines can choose it in the Codex command-line tool and the new editor extension. OpenAI also says it plans to offer the model to developers through its API soon.

What OpenAI released

The announcement bundles a new model with several upgrades to the Codex product, which two weeks ago became one experience tied to your ChatGPT account:

Piece	What is new
GPT-5-Codex	A GPT-5 trained further on real engineering work: building projects, adding features and tests, debugging, large refactors and reviews
Codex CLI	Rebuilt around agent workflows; accepts screenshots and diagrams, keeps a to-do list, can search the web and connect to tools through MCP
IDE extension	Brings Codex into VS Code, Cursor and other editors built on VS Code, and hands work back and forth with the cloud
Codex cloud	Faster task start-up, automatic environment setup, and a browser it can open to check what it built
Code review	Reviews GitHub pull requests on its own and can fix what it finds

A refactor is a reorganization of existing code that keeps its behavior the same. MCP, the Model Context Protocol, is a common standard for plugging AI into outside tools and data.

Codex is part of five ChatGPT plans: Plus, Pro, Business, Edu and Enterprise. OpenAI says the Plus, Edu and Business plans cover a few focused coding sessions a week, while Pro can support a full week of work across several projects.

How it adapts its effort

The most practical change is how GPT-5-Codex decides how hard to think. Reasoning models produce hidden "thinking" before they answer, and that thinking costs time and tokens, the small chunks of text a model reads and writes.

GPT-5-Codex scales that thinking to the job. On OpenAI's own employee traffic, for the lightest tenth of requests it used 93.7% fewer tokens than GPT-5. For the heaviest tenth it went the other way, spending about twice as long reasoning, editing and testing. In OpenAI's testing it worked independently for more than 7 hours on big, complex tasks, fixing test failures along the way until the work was done.

Think of a good colleague: a quick question gets a quick answer, while a request to restructure the billing system gets a careful plan, many small steps and repeated checks.

OpenAI's system card addendum explains how it was trained. Like the earlier codex-1 model, it learned through reinforcement learning, practice with rewards, on real coding tasks in many environments. The goals were code that looks like a careful human wrote it, close attention to instructions, and running tests repeatedly until they pass.

On scores, GIGAZINE's coverage of OpenAI's charts reports that GPT-5-Codex did slightly better than GPT-5 on SWE-bench, a test built from real bug fixes, and much better on OpenAI's refactoring test: 51.3% of tasks completed versus 33.9% for GPT-5.

Code review on GitHub

The second big addition is review. A pull request is a proposed change to a codebase that teammates check before it is merged. Once you turn Codex on for a GitHub repository, it reviews pull requests automatically when they move from draft to ready, and posts its analysis.

Unlike a simple style checker, it compares what the pull request says it does with the actual changes, reasons across the whole codebase, and runs the code and tests to confirm behavior. OpenAI says that internally Codex now reviews most of its own pull requests and catches hundreds of issues a day. You can also ask for a review directly in a comment:

Prompt

@codex review for security vulnerabilities

If the review suggests fixes, you can ask Codex in the same thread to make them. OpenAI is clear that this should be an extra reviewer, not a replacement for a human one.

Try it: Pick a small, finished pull request in one of your repositories, ask for an @codex review, and compare its comments with what a teammate noticed. Note which issues it found that a person missed, and which comments you would ignore.

Safety and limits

By default Codex runs in a sandbox, a sealed-off environment, with network access turned off, both on your computer and in the cloud. That limits the damage a mistake can do and lowers the risk of prompt injection, instructions hidden in web pages or files that try to hijack the agent.

The CLI now has three approval levels: read-only, where you approve every action; auto, which can work freely inside your project folder but asks before going outside it; and full access, which can read anywhere and use the network. Each task comes with logs, test results and citations so you can check the work. As with GPT-5, OpenAI treats the model as highly capable in biology and chemistry and applies extra safeguards there.

OpenAI recommends GPT-5-Codex only for coding work in Codex or similar agent tools; for general questions, GPT-5 remains the model to use.

How can you start using GPT-5-Codex?

If you have an eligible ChatGPT plan, the simplest route is Codex on the web: connect a GitHub repository and give it a small, well-defined task, such as adding tests for one file. GPT-5-Codex is already the default there, and you can review its changes before anything is merged.

To work on your own computer, install the Codex CLI (npm i -g @openai/codex) or the IDE extension in VS Code or Cursor, sign in with your ChatGPT account and select GPT-5-Codex. Start in read-only or auto mode so you see each step, and write a short AGENTS.md file in your project describing how to run tests and what style to follow; OpenAI says the new model follows these instructions more closely.

Finally, turn on code review for one repository and watch how it behaves for a week before relying on it more widely.