I’ve been using Codex pretty heavily as part of my development work, and lately I’ve been running into a problem that I think could be address that problem. I’m a Plus subscriber, and I understand that I’m paying $20 a month and not $200 a month and I don’t expect the same amount of usage that somebody on a Pro plan gets. And I’m not asking OpenAI to simply raise my limits but what I would really like is some way to manage the usage I already have.
My particular problem is that I tend to give Codex fairly substantial prompts. I’m not a software engineer, so I use Codex to carry out fairly involved development tasks for me, often involving multiple files, testing, verification, documentation, etc.
What has been happening lately is that I can sometimes use up my available Codex time surprisingly quickly. The bigger problem, though, isn’t simply that I hit the limit. It’s that neither Codex nor I seem to know before starting a task whether there is enough usage left to actually finish it. That can leave me in a much worse position than simply having to stop working. For example, Codex may start a substantial task, modify a number of files, get most of the way through the implementation, and then run out of usage before it can finish testing and verifying what it did. When my usage becomes available again, I then have to spend part of that new allotment figuring out where it stopped, whether everything is in a safe state, what still needs to be tested, and how to resume the work.
So I’ve been thinking that what would really help is some kind of resource preflight before Codex accepts a large task. It doesn’t have to be perfectly accurate. Even a rough estimate would be enormously useful. Something along the lines of:
“Based on the size of this task and your remaining usage, I believe I have enough capacity to complete it.”
Or:
“This appears likely to require more usage than you currently have available. I can either perform the first part of the task and stop at a safe checkpoint, reduce the scope, or you can wait until your usage resets.”
Even better would be if Codex could roughly break a larger task into stages. For example, if it could tell me that it probably has enough capacity to perform the implementation but may not have enough remaining to complete testing and verification, I could choose not to start. Or perhaps I could tell it to complete only the first phase and leave the repository in a known, documented state.
I’d also really like to see a clearer indication of how much Codex usage I actually have remaining and when it resets. A percentage or even an approximate gauge would be fine. I’m not looking for accounting down to the last token. I just want enough visibility to make a decision before asking Codex to undertake something expensive.
The other feature I think would be extremely useful is a kind of safe-shutdown behavior.
If Codex knows it’s getting close to whatever usage boundary applies, I’d much rather have it stop beginning new work and instead use what remains to get to a safe checkpoint: finish the current operation if possible, run whatever verification it reasonably can, tell me exactly what it completed, tell me what remains, and leave me a clean continuation point. That would actually make the available compute more valuable, because I wouldn’t have to spend part of the next usage period reconstructing what happened during the previous one.
I realize estimating the cost of an AI task in advance is probably not simple. Context size, model choice, reasoning, tool use and unexpected problems all affect it. I wouldn’t expect the estimate to be exact. But even something like “low / medium / high expected usage,” or a percentage range with a confidence level, would be far better than going into a large Codex task completely blind.
I really like Codex. That’s actually why I’m raising this. It has become important enough to the way I’m developing things that running unexpectedly into the usage boundary has become genuinely disruptive. I can’t justify spending $200 a month on Pro at this stage, and I suspect there are a fair number of Plus users in a similar position. We’re willing to work within a limited amount of compute. We just need better tools for managing that compute. I’m not asking for unlimited Codex usage. I’m asking for Codex to help me understand whether it has enough gas in the tank before we start driving.