So the core concept of your message has been addressed.
You’ve been advised on what you can do to make the codex usage last longer.
Have a good day.
I’ve noticed the 5h limit seems incredibly low compared to any other week. I’ve been using Codex since it’s launch in February and am very aware of the 2x limits that just expired. Even so, I was fairly happy with how much usage I got out of a $20 plan. Today, asking to summarize a single method to verify if the pre-multiplication was working as intended used 6% of my 5h usage. It didn’t code anything, or read large files, or even work in a giant repo.
I followed up with a simple prompt to modify the single method, and it took another 5% of my usage out.
I’m not sure if it’s a bug or what, but this is the first time I’m logging into the forums to even say anything, as it’s really bad today. Hopefully the OAI team can figure out what’s going on, as at this level, I’ll get about 10-15 minutes of usage before the 5h timer is up even in a small single app project with a dozen class files.
I already know that workflow. I’ve used Codex that way too, and I still do when it makes sense.
What you keep framing as some deep insight is really just a basic workaround. A lot of people understood that from day one. The difference is, not everyone is using Codex for tiny edits, simple scaffolding, or toy-level tasks. Some of us are using it for heavier technical work, where sustained iteration matters and the current limits become a real constraint.
So no, this is not about me being vague or not knowing how to prompt. It’s about you assuming your lightweight usage pattern applies to workloads where it clearly doesn’t.
And the patronizing tone doesn’t help your argument. It just makes a pretty shallow point sound more important than it is. I know the workaround already. I’m saying it’s not enough for the kind of work I’m actually doing. If a product only feels “good” when people narrow their usage and push the rest somewhere else, that’s a product limitation, not user error.
Ok boss.
Ive used it nonstop for days, almost every waking moment on 2 pro accounts an not had any issue.
planning seems to be key man.
So I can agree with that, my experience is also with GPK 5.4, it doesn’t matter if you work with low or high, it doesn’t matter, at just under an hour, 5 hours are over and whether you work with low or high reasoning, when the model says I don’t need the power, it doesn’t matter if it’s longer or shorter, it doesn’t make a difference. In the real software development area, just a few minutes, if you differ from low to high reasoning, so you can always run on high, I’ll probably run the days on extra high and see if extra high also lasts about 45 to 60 minutes. And the other recommendation, for example, to switch to 5.4 mini is also pretty banana, because 5.4 just produces less cool results and you need longer and it just doesn’t work for you, so the only option is to switch to 5.3.
5.3 doesn’t make a difference. My metrics are with 5.3 high already, trying to get the most effective usage. Anything less than 5.3 high makes so many more errors that it normally does that I need either to spend a half an hour after each small task to fix it by hand or burn more limits for the same task – and still having to fix issues manually afterwards anyhow 99% of the time (that’s on C++ tasks, YMMV with simpler workloads).
Hey people, did anyone find a good alternative? Until now, Deepseek seems to be the best one in terms of balance cost/quality/usage on the coding side but they don’t offer yet anything similar to the custom GPTs and there’s no connectors (Google Drive, Hubspot ecc). I’m trying to make things work with this new Codex rate limits but it’s too castrating and time wasting even following the usage suggestions and I need to get back to full operativity with an alternative by May 9th, that leaves me go ahead at the same rhythms as before and doesn’t make me feel like I’m wasting my money. A good deal basically. Any good suggestion will be really appreciated.
Your noting of when and how to use “Hide Details” is appreciated.
One side effect that many may not know with “Hide Details” is that the content in the hidden section is not seen by crawlers as it is not there in the HTML until the button is clicked.
I use “Hide Details” often to post things that are of value but that I know are wrong and should not be indexed, or for other various reasons where I want the content to be found by humans but not via a search engine.
I really do appreciate every little tip about the UI here…
That was valuable knowledge,
It was just one of those features that seemed simple enough that I assumed.
@EricGT look at your most recent post in this thread, which is 2 posts up from here.
Click on replies, and as you were technically only replied to a single time in that post… I was curious if the second reply created notification spam?
It’s a concern about notification hygiene going forward, which…
I’ve sorely abused you with today already ^.^
Most people don’t care about leaving stale information in their posts or even misinformation around on the internet.
Generally, when I’m updated or corrected, I push it all to ‘origin’ so to speak.
Not many people bother to.
But with the direction that Ai is going, that sort of bad internet hygiene makes for noisy signals for AI to sift through… can cause drift… waste tokens…
Dear God I sound like a techno-hippie now
Thank you for your patience today ^.^
Is there any investigation for GPT-5.5 (in high mode for example)? I also found that just compressing context roughly consumed 8 % or so of the 5 hour window… That is really a lot, just for that task.
I’ve been using Codex since the very beginning, but I haven’t done much over the past three weeks—just some minor tasks in my repositories—so I didn’t even notice the change.
Today I tackled another big task, and it makes me angry that a single prompt used up 50% of my 5-hour limit and about 20% of my weekly limit. With the current Pro subscription, that’s a huge joke.
After my third big prompt, I hit the 5-hour limit and was completely surprised by it… I guess I’ll go back to Co-Pilot and Claude or something, because ChatGPT with these limitations is a joke and not worth a single dollar for the tasks I do—and other AIs can handle them just as well.
What a disappointment from OpenAI and Sam Altman, but the greedy and corrupt ones always abuse their power.
Out of interest is that the $100 or $200 Pro plan? The latter offers 4x over the $100 plan. (see Introducing New $100/month Pro Tier - #9 by merefield)
Which model were you using?
How can you justify such a statement without considering the business cost of providing you with that inference?
You have been enjoying subsidised inference and gradually improving peak capability.
It should be no surprise that service levels and prices would eventually have to be adjusted, especially for the bigger models that perform a lot of reasoning and use up a lot of limited compute (whilst delivering significantly improved capabilities)
I can see people might get frustrated, but that’s no reason to make apparently baseless accusations of corruption?