OpenAI introduces a series of improvements to Codex designed to turn it into a more integrated and practical development companion. Can you imagine delegating repetitive tasks, seeing clear diffs and reviewing pull requests from your phone? That's exactly what this update aims for, and yes, it comes with controls so you don't lose the reins. (openai.com)
What's changing in Codex
The updates are concrete and focused on fitting into your workflow: Codex is no longer just a demo inside ChatGPT — now it’s available from the terminal, IDE extensions like VS Code, web integrations and via the CLI. You can start tasks, generate patches, run tests and push PRs without leaving the environment you already use. (openai.com)
Also, a set of technical capabilities was added to improve the experience and operational safety:
- Internet access during task execution, optional and controllable by domain and HTTP method. This lets you install dependencies or run tests that need external resources, but it’s turned off by default. (help.openai.com)
- Support for updating existing pull requests when Codex is assisting a task. Great for iterating without opening a new PR every time. (help.openai.com)
- Voice dictation to create tasks and UI improvements for reviewing diffs and applying patches. They also increased time and size limits for setup scripts and diffs on Pro and Team plans. (help.openai.com)
- Generation of multiple simultaneous responses (Best of N) so you can explore several quick solutions. (help.openai.com)
What does this mean for you as a developer or team?
First: real productivity. If you work on repetitive features, refactors or tests, Codex can propose changes, run the tests and prepare a PR ready for review. Think about delegating the first draft of a fix and reviewing it in five minutes instead of writing everything from scratch.
Second: integration into your flow. Do you use VS Code or GitHub? The extension and improved repo connections let the agent act inside the processes you already have, without pushing you to new APIs. That lowers friction for teams that want to experiment with AI without reconfiguring everything. (openai.com)
Third: mobility. You can start and review tasks from the ChatGPT mobile app, which makes quick reviews outside the desk much easier. (help.openai.com)
Practical risks and how to mitigate them
Sounds perfect and a bit scary? Totally normal. There are concrete points to watch:
- Internet access: useful but it can expose dependencies or allow unwanted calls. Enable access only for controlled environments and define a whitelist of domains. (help.openai.com)
- Mandatory human review: don’t accept PRs automatically. Use Codex to speed up work and keep final approval with a reviewer. This prevents silent bugs or unintended changes from reaching production.
- Traceability: check the diffs and commit messages Codex generates. Ask for small, verifiable tasks so you can audit its behavior.
Concrete recommendation: enable internet only in staging environments and for specific domains, and set up alerts when Codex attempts to install new packages.
How to get started today
- If you have ChatGPT Plus, Pro, Team, Edu or Enterprise, Codex is already available to try in your account. Check the official docs and the changelog to learn limits and permissions per plan. (openai.com)
- Start with small tasks: create tests, rename functions or generate a setup script. Review the results and tweak prompts before scaling usage.
What you should ask now
Do you want Codex to act only as an assistant that proposes changes, or should it also execute and create PRs automatically? Does your team need to review security policies before granting internet access? These choices affect how you configure permissions and workflows.
OpenAI is moving Codex from an experimental tool to an agent integrated into daily development. That eases a lot of repetitive work, but the key is setting clear limits and keeping human review. Ready to try it and design the guardrails for your team?