I am not looking for another coding subscription.
Cursor is already where I work. I have also tried Claude Code and revisited GitHub Copilot. The pattern is familiar: I stay curious, I stay impressed in places, and I still open Cursor the next morning.
Then I noticed that the ChatGPT Plus subscription I already pay for personally may include access to OpenAI’s Codex coding agent. That changed the question. It was no longer “should I buy another tool?” It became “should I give Codex a proper try with something that already exists?”
I have not used Codex recently. This is the test I want to run, not a hands-on review.
Codex is not “an OpenAI model inside Cursor”
That distinction matters.
Using an OpenAI model through Cursor is one thing. Codex, as OpenAI describes it, is a coding agent product in its own right: surfaces such as the web, CLI and IDE extension, signed in with a ChatGPT account. OpenAI’s help centre says Codex is included with ChatGPT Plus, with usage limits rather than unlimited free rein. Plus is described as enough for a few focused coding sessions each week, not an open tap.
I already have ChatGPT Plus personally. That is not my work account, and personal subscription access is not automatic permission to point any agent at every codebase I touch. “I can sign in” is not the same as “this is fine for everything.”
Let’s go big or go home
My instinct is not to ask Codex to rename a button and declare the experiment finished.
If I am going to learn anything useful, I want a substantial real-world test. Go big or go home.
The candidate is an existing Streamlit application used for internal finance tracking and profit-and-loss calculations. It is the same application I wrote about when Cursor helped redesign its tables and dialogs. The app already exists. Codex has not built it, audited it or maintained it yet.
I would start by asking Codex to understand and audit the structure, then explore how it could help maintain and extend the thing. It is a living application with history, not a blank folder waiting for a demo.
Why understanding matters more than a clever first pass
This is not boilerplate assembled from a default SaaS template.
The application has an established look and feel, context documents, rules and implementation decisions. I do not want a coding agent to “improve” it into a generic product shell. Conventional design patterns are not automatically bad. Replacing deliberate choices without understanding why they are there is.
So the real test is whether Codex can respect what already exists while making a requested change. Saying “I understand” is a starting claim. Implementation and review are what would show whether that claim holds.
The workflow I want to inspect
Around that test, I also want to see how Codex fits the surrounding work: Git, project instructions and context, reusable skills, hooks and checks, deployment to a testing environment, and ongoing maintenance.
OpenAI’s Codex docs describe project guidance through files such as AGENTS.md, reusable skills, and hooks that can run around the agent loop. There is also an IDE extension path. Those are documented capabilities I would evaluate. Formats and behaviour differ from Cursor, so I would have to adapt anything I cared about.
I am interested. I have not already wired it.
The November module idea
The practical shape of the experiment is simple enough.
I would create a separate branch. A branch helps me review and isolate changes. It is not a security boundary and it does not stop mistakes on its own.
I would ask Codex to understand the application and its rules. If it reported that it understood, I would test that claim by asking it to implement a real new module. I have two new modules planned for November and I am considering giving Codex one of them.
I have not run that test. I have not reviewed Codex output for this app. I have not deployed its changes. I have not chosen it over Cursor.
The open question
I am willing to explore Codex without abandoning Cursor.
The useful comparison is not which logo I like more. It is whether Codex earns a role in the process I already have:
Can it follow the rules, preserve the application’s identity and complete a substantial module in a form I can review?
I do not know yet. That is exactly why a tiny demo would waste the chance. If I am going to try Codex at all, I want the answer to come from work that matters.
