Up until around two weeks ago, I was genuinely impressed with Codex. I thought it was expensive in terms of credits, but the results were good enough that I was happy to use it.
Over the last couple of weeks, though, my experience has changed dramatically.
Today I gave it a fairly straightforward website task. The first result was actually excellent. I then asked for a few small refinements, and from there it became an endless sequence of supposed fixes that either made no visible difference, addressed the wrong thing, or introduced new problems.
Each time it would confidently tell me it had identified and corrected the issue, only to acknowledge afterwards that it hadn’t actually addressed the requirement.
The worst part is the credit usage. I started this morning at 100% usage remaining. During this one relatively small task it dropped to 10%, with much of that apparently consumed by repeated attempts to correct its own unsuccessful changes.
At that point I stopped it.
I’m posting this because the difference compared with my experience a couple of weeks ago is enormous. Codex went from something I was praising and actively using to something I’m now reluctant to touch because a simple task can consume most of my allowance.
If OpenAI staff monitor this forum, please look into both the recent reliability of Codex and the amount of usage being consumed when it repeatedly attempts to correct its own work.
Have you tried asking the model to provide an overview of your codebase and rank it based on healthiness?
100% → 10% from a single task is a serious problem. If you cannot understand what the model is doing, then it’s time to fix up your workspace.
This is the most telling part. I think most people who jump into a coding project with a model all suffer from the same fate, because even coders themselves suffer it!
Starting is always the easiest part. Each iteration you make - especially if you haven’t fully conceptualized the project - can add technical debt to the project. It doesn’t help that the model is always trying to implement production-ready code in a codebase that’s meant to iterate fast and break loudly.
It’s fine if you don’t code, or don’t care to, but it’s not fine if you blow your usage budget and have no idea what happened.
Ask the model to review your codebase for healthiness
Send the Codex session to ChatGPT and ask if anything strange is happening, and where all its time is being used up
For anyone having these issues, try recording what Codex is doing with the --json flag. Here is an example I commonly use when iterating with ChatGPT:
I’m also seeing unusually high usage on my end, and the difference compared with before is significant—at least 4–5 times higher. The gap is large enough that I’m confident this isn’t due to a change in how I’m using it.
It wasn’t anywhere near this bad just two weeks ago. These are simple, light tasks too — things like changing a single module on a webpage.
Codex now seems to get stuck working on the same task for 20 minutes or more, consuming a significant amount of usage in the process. It then reports that the task has been completed, but when I check the result, either nothing has changed or the requested work hasn’t actually been done.
The frustrating part is paying for all that repeated processing and ending up with nothing usable to show for it. Something seems to have changed quite dramatically in the last couple of weeks. I won’t be buying more credits.
Guys, it takes a single command when running Codex and then passing the file over to ChatGPT for diagnostics.
There is just way too many unknowns to even consider trying to diagnose your issues.
All I can say is that increased complexity and model run-time is expected each time you iterate in your codebase. Even programmers have to refactor numerous times as truths become fallacies and paradigms start causing friction
You have completely lost control at this point. Are you using Git? Run a diff on the change and see what’s happening. It’s more than likely that the model is performing what you ask, but what you see and what the model is dealing with is completely disconnected.
ChatGPT is extremely capable of understanding what’s going on, you just need to be capable enough to atleast provide sufficient information.
Has anyone noticed that account usage limits seem to have been reduced significantly recently, perhaps because of issues with the risk-control system or related safeguards? For example, the weekly quota for X20 Pro may now be only a little over 200.