Just after I renewed my ChatGPT Pro membership and felt that it was not worth it to just buy the usage of o3 and DeepResearch, OpenAI released the ChatGPT Codex feature.

The Codex function is currently exclusive to Pro members, and I don’t know when it will be available for download.

Basic Information

Official introduction page: https://openai.com/index/introducing-codex/

Codex's OpenAI official press conference was almost like watching nothing. There wasn't much substantive information. Basically, Codex can do a lot, it's very good, and it uses end-to-end RL training. (Many guests seemed to have difficulty speaking. However, it's too difficult to ask developers and researchers to do it at the PR level. After all, proficiency requires rehearsal. Let them do the work at this time.)

Codex is a cloud-based AI coding agent, which looks very much like a basic version of Devin. It does not include rendering front-end pages and understanding, browsing web pages, etc., but its pure code functions are in line with my expectations of Devin. However, its workspace cannot be persisted in a conversation session. And Codex can be accessed on mobile phones, enjoying the benefits of cloud execution.

The entire workflow of Codex depends on Github, including git repo hosting, PR, etc.

During the conversation, Codex can output reference markers to the original code.

Latent Space put on a Codex podcast titled "ChatGPT Codex: The Missing Manual", which talked about some internal design considerations. However, it didn't mention much about the boundaries of capabilities.

ChatGPT Codex: The Missing Manualhttps://www.youtube.com/watch?v=LIHP4BqwSw0

Personal experience

In some small tasks, I feel that Codex meets my expectations and feels better than Cursor Agent mode. However, the whole process still seems to have some problems caused by the instability of the environment.

But one limitation is that Codex can only handle one branch per task. I studied the environment configuration for a long time, but I couldn't manually specify it to pull multiple branches. I don't know if it was intentional or for some other reason. The overall feeling is that the environment configuration is not very good. Internet access will be cut off when the agent is running , and can only be accessed during the initialization of the environment. The reason is for security reasons. When I was testing, I couldn't access it during initialization either. I don't know if it's a bug.

But overall, the Codex product based on RFT and o3 addition makes me very interested in using it. What I lack now is a product that is smart enough.

Personal comments

There are many new code tools recently. Anthropic released Claude Code, OpenAI released Codex CLI and Codex (cloud).

Among these tools, Codex is probably the only one that uses RFT/end-to-end RL. OpenAI’s first RFT product, Deep Research, has already become famous, so I am also looking forward to Codex this time.

At present, I feel that Codex runs very fast. The official limit of running time is 1 hour. The official said on the Latent Space podcast that the time for difficult tasks is 30 minutes. The current task concurrency limit is 60 per hour.

Now, OpenAI's Deep Research product line has three versions: the full-featured Deep Research, Deep Research Lite, and o3+Search. The last o3+Search is already a lightweight Deep Research, which is very useful. The search rounds and exploration time are not small, and the actual cost should be significantly higher than the simpler solution.

Some technical topics

Codex makes extensive use of traditional Linux ecosystem tools such as grep, nl, sed, etc. This may be different from what many people imagine. Combined with the voice of the Claude Code team, AI Coding Agent's extensive use of existing coding tools is a more appropriate path (higher ROI).