OpenAI upgrades Codex with GPT-5 for programming

OpenAI has updated Codex with GPT-5-Codex, a model specialized in programming tasks that can work from the terminal, code editor, GitHub, the web, and the ChatGPT app for iPhone.

The idea is for it to do more than answer questions about code. Codex can modify a project, run tests, fix errors, and continue working independently on long-running tasks.

A model that adjusts its effort

GPT-5-Codex is trained for common software engineering work: creating projects from scratch, adding features, generating tests, debugging problems, reorganizing large parts of an application, and reviewing changes before they reach production.

The model adjusts how much time it spends on each task. For a small request, it should respond quickly. For a complex reorganization, it can analyze the project, try different solutions, and repeat the process until the tests pass.

In OpenAI's tests, it worked independently for more than seven hours on large and complex tasks. The company also says that, among the 10% of interactions that used the fewest tokens, GPT-5-Codex used 93.7% fewer tokens than GPT-5. In the 10% of most demanding tasks, however, it spent twice as long reasoning, editing, and testing the code.

These figures come from OpenAI's internal traffic, not an independent measurement. They help explain the model's goal: to be fast when the task is simple and do more work when the problem requires it.

Code review and teamwork

One of its main features is automatic code review. Codex can go through a project, study its dependencies, run the code, and check the tests for serious problems, rather than simply looking for suspicious patterns in a file.

When connected to a GitHub repository, it can automatically review pull requests when they move from draft to ready for review. You can also request a specific review by mentioning @codex review, or give it a specific goal, such as looking for security vulnerabilities or outdated dependencies.

If it finds a problem, you can continue the conversation in the same thread and ask it to prepare a fix. OpenAI says Codex already reviews most of its own pull requests and detects hundreds of problems every day, often before a human review begins.

That does not make the agent a replacement for the team. OpenAI recommends reviewing its work before modifying or publishing code and using its reviews as an additional layer of security.

Available wherever you already program

Codex brings several ways of working together under the same ChatGPT account. You can start a task in the cloud and continue it on your computer without losing context, or move from GitHub to your editor to finish the changes locally.

The update includes:

• A new version of Codex CLI, the tool for using it from the terminal.
• An extension for VS Code, Cursor, and other VS Code-based editors.
• GitHub integration for delegating tasks and reviewing pull requests.
• Image support in the terminal, editor, and cloud for sharing screenshots, designs, or diagrams.
• Web search and connections to external systems through MCP.
• Task lists for tracking progress on longer jobs.

In the editor, Codex can use open files and selected code as context. This lets you write shorter instructions. In the cloud, it can also open a browser, visually check an interface, and attach a screenshot of the result to the task or pull request.

More speed, but with controlled permissions

OpenAI says that caching execution environments reduced the median time by 90% for completing new tasks and follow-ups. Codex can also detect a project's usual setup scripts and prepare its own environment.

By default, it runs in an isolated environment with no internet access. Permissions can be expanded, but each level carries a different risk: from reading files and requesting authorization for every action to allowing full access, command execution, and network access.

This matters because an agent that can execute commands can also cause damage if it receives malicious instructions from a file, dependency, or web page. The company recommends reviewing permissions, limiting trusted domains, and checking the logs, citations, and test results Codex provides.

Who can use it

Codex is included in the ChatGPT Plus, Pro, Business, Edu, and Enterprise plans, with limits that vary by plan. Business teams can buy additional credits, while Enterprise uses a shared pool based on usage.

On September 23, 2025, OpenAI announced that GPT-5-Codex was also available to developers using Codex with an API key. The model is offered at the same price as GPT-5 and, in the API, is available only through the Responses API. Its internal version will be updated regularly.

For you, the main change is practical: Codex can handle more real work, not just suggest lines of code. It can prepare a feature, run the tests, and leave you with a proposal to review. What remains essential is deciding what it should do, checking that it did it correctly, and keeping control over its permissions.