Introducing the New Codex for (almost) everything

Announcements
,
1

Codex is getting a major update. Already used by millions of developers, it is now expanding beyond coding to support a broader range of work. The focus remains clear: stronger developer workflows, better integrations, and less friction across projects.

Persistent Context and Intelligent Automations, Coming Soon

OpenAI is releasing a preview of memory. Codex will also support future work scheduling and more proactive help with ongoing projects. These personalization features, including memory and context-aware suggestions, are intended to improve productivity across multiple sessions and longer-term workflows.

Improved Developer Productivity

This update includes several additions aimed directly at developer use cases:

Full lifecycle coding support: Codex can now help with GitHub pull request comments, supports multiple terminal tabs, offers richer file previews including PDFs, spreadsheets, and presentations, and adds alpha support for SSH connections to remote development environments.

Browser-based workflows: An integrated browser lets you give precise instructions directly on web pages, which can make frontend development and game iteration faster and more intuitive.

Project summary pane: You can quickly review plans, sources, outputs, and artifacts in one place to stay organized.

More Integrations and Plugins

Codex now supports more than 90 new plugins and integrations for tools many teams already use, including Jira, GitLab Issues, Microsoft Suite, CI/CD pipelines, and databases. This makes Codex more useful as a central workspace across development and productivity workflows, while reducing unnecessary switching between tools.

Native macOS App Interaction

Starting on macOS, Codex can now interact directly with desktop applications. This opens up a new layer of automation for tasks that would otherwise need to be handled manually.

Built-in Image Generation for Multimodal Workflows

Codex now includes native image generation, making it easier to create and iterate on product mockups, concept visuals, UI prototypes, and game designs without leaving the workspace. This expands what you can do in one place and adds useful support for visual workflows.

Availability and Rollout

The updated Codex begins rolling out today for Codex desktop app users signed in with ChatGPT. Desktop app interaction, initially on macOS only, and browser integration will reach users in the EU and UK shortly after. Personalization features, including memory and proactive context-aware suggestions, will be available soon for Enterprise, Edu, and EU/UK users.

Codex is no longer just about writing code. This update moves it further toward being a practical tool for handling the broader workflow around development and day-to-day project work.

3

This is super! I am such a big fan (and user) of Codex, it’s great to see it expand its horizons.

4

Codex with imagen included is a lot of fun, it just makes everything look so much better! Nice to see it becoming an all purpose agent, there is a lot of possibilities beyond just coding.

6

Now we got chats in Codex. How the limits are distributed? I opened chat and I think it started in Codex Win App and I think, it started to consume Codex limits, although there’s no limit info underneath (like in local project).

7

Can someone explain to OpenAI that especially among developers Windows is the dominant operating system and not MacOS? Making a coding app for Mac seems like a bad decision IMO.

Group Windows macOS
Developers (professional) ~48% ~31–33%
Developers (personal use) ~59% ~32%
General users (desktop/laptop worldwide) ~60–71% ~14–16%
All devices worldwide (incl. mobile) ~26–30% ~2–4%
8

What?

33% is a huge market and anecdotally having worked as a professional freelancer for nearly 10 years my experience is that most other freelancers/digital nomads I meet in person are carrying a MBP. (NB I’ve been in IT a lot longer :sweat_smile: and have been on many floors that are a sea of Windows machines)

I’m sure OAI’s aim is to support the broad feature set across both platforms, if not Linux too … we just need to be patient.

9

I think you’re close to the real reason they prioritize macOS. They likely develop first on Linux for internal testing. Then they move to macOS, since it’s a Unix-based system, which makes porting relatively straightforward. Finally, they support Windows. Although WSL2 is a great tool, not all Windows developers are familiar with it, so they still need to provide native Windows support—which is more challenging when starting from a Linux environment.

Can anyone confirm whether this analysis is correct?

Franz

10

Imho if someone is looking at trying to get into professional development work and their primary machine is windows, they should learn how to use WSL. It is indeed fantastic.

11

I’ve been waiting for some link between Codex and Chatgpt for months. I find it amazing that this hasn’t happened. I run around 20 Pro chats in Atlas, and desperately need to link them to codex directly. Otherwise it just gets mind boggling. It feels like not many others use Pro this way, which is really mysterious. It saves SO MUCH TIME to have pro do things correctly the first time (yes, using feedback from other models), rather than have codex implement a plan its created that misses some key consideration.

12

Have you thought about using Codex CLI? You can log into Codex CLI with your ChatGPT account and have these chats directly in Codex?

The only flaw in that approach atm, imho, is that session management in Codex CLI is a little on the clunky side compared to the way you can so easily navigate chats in the ChatGPT web UI and there are probably ways to improve it that they haven’t thought about yet/got around to.

13

The only reason I want to connect chatgpt to codex or codex cli is so that chatgpt Pro has full visibility of a repo. It would obviously be ideal if codex cli had pro in it, but it doesn’t. It may be available in the cli sidebar, but i believe that runs through the API so it would be outside the Pro plan and end up possibly costing thousands per month. There’s probably a reason why they aren’t enabling pro in terminal/cli/vscode/codex which may have to do with the ‘many worlds’ branching overwhelming things?? But all im really looking for is not for Pro to interact, but that it be able to read a full codebase, not just a couple of files.

14

Codex CLI gives access to 5.4 at xhigh via your ChatGPT account which is seriously impressive.

15

Hi @barker.edward!

I also use the Pro model in ChatGPT to challenge Codex output. Besides helping to learn where the Codex models currently have limitations, that kind of feedback loop is absolutely the right approach!

Have you tried using custom GPTs or ChatGPT Apps in developer mode? You can store partial model replies through actions or MCP tooling, then connect that with local Codex sessions through skills.

One thing to keep in mind: automatic extraction of ChatGPT outputs is still against the Terms of Service and could cause issues. But if you set things up so that only the code or plan review is stored externally, you can build your own solution fairly quickly.

Hope this helps!

Ps. In fact, I believe the latest version of the GitHub App for ChatGPT has write permissions. This could be the fastest way to set this up.

16

(Rhetorical question ahead) :slightly_smiling_face:
Have you joined in on this discussion?

I do not know if they still need help but they have not taken down the discussion yet.

17

Honestly, the new features are not the main issue. The limits are.

A coding assistant is only useful if it can survive normal iterative development. Right now, the limit model feels too restrictive for real work. It breaks momentum, discourages experimentation, and turns ordinary coding flow into quota management.

That makes all the “new Codex for almost everything” messaging hard to take seriously. If using the product normally burns through limits too fast, then it is not built for “almost everything.” It is built for small, cautious usage bursts.

OpenAI needs to fix the limit system first. Otherwise the product keeps improving on paper while becoming less usable in practice.

18

I really hope Codex can work with local files on disk without using 250% CPU and making the computer burn up, the bug been around for a long time now… (codex with VScode)

19

I agree with your sentiment. Normally I spend only $20./month for ChatGPT as a Plus user and it has helped me develop multiple sophisticated AI driven applications. But when I tried Codex the other day, I spent $60. on API fees in that one day alone. But I’m convinced it was worth it, because it found a bug in my provider switching code that it fixed and now I can simply switch providers within my app if OpenAI or Anthropic cut me off due to limits. This is OK, because I also run a local LLM that is free, even if not as good as either Claude Code or Codex. But Codex deserves the credit because in a single day of using Codex, I feel that I made 6 months of progress. Three new features were integrated into my projects and Codex got it right the first time each time. I’m very impressed with the capabilities of Codex, but I can’t afford $60./day for API inference fees. Thanks for bringing this out. Regards, Glendon Gross

20

You could probably get better value by using GPT-5.4 through Copilot for around $10/month instead of burning $60 in a single day.

That is the real issue with Codex right now. It used to feel more efficient than Copilot, but with the current limits, it has clearly fallen behind. The model may still be strong, but the product experience is not.

Instead of focusing on development, debugging, and catching real issues, you start worrying about whether your limit is about to run out and when your workflow will get cut off again. That is not productivity. That is quota anxiety.

21

I agree 100 percent. I think the reason that it burns so many tokens is that it spins up multiple agents to accomplish the tasks you give it. This is very effective, but has the effect of hiding the costs from the user. Sticker shock may result.