I consider the way Codex usage limits are presented to be misleading.
The product shows a “5 hours” usage limit/window, but in real work on a serious repository this allowance can be exhausted in about one hour, or after only a few tasks. For a normal customer, “5 hours” clearly sounds like available working time. In reality, it behaves more like a usage/compute/token budget that can disappear extremely fast depending on repository size, context size, model choice, task complexity, and token consumption.
This is not transparent enough.
If the limit is actually based on credits, compute, tokens, or internal usage cost, then the interface should not present it primarily as “hours” without a clear warning and a detailed breakdown of what consumes the allowance. Showing “5 hours” while the practical usable time can be around one hour feels misleading.
I am extremely disappointed with this experience.
There are different speed options, but from a user perspective the displayed “5 hours” still creates the expectation of usable working time. If a fast or top model can burn through the allowance after only a few tasks, the UI must make that obvious before the user pays. Otherwise, OpenAI could theoretically display any large number of “hours” while the real usable work time is much smaller. That is not a fair way to present limits to paying users.
In my case, Codex feels severely restricted for real project work. I paid expecting a practical coding assistant, but instead I get a system that can consume the whole allowance very quickly and then leaves me unable to continue meaningful work.
This is especially frustrating because the top model may be useful for large and complex tasks, but it can also consume the entire allowance too quickly. After that, the user is stuck. There should be a better fallback for paid users: for example, allow continued daily use with a lower model after the top-model allowance is exhausted, so users can still perform smaller routine tasks while waiting for the main limit to reset. Right now, the experience feels like I paid for “5 hours” of Codex, but in practice I may get only about an hour or a few serious runs before the system becomes unusable for real work.I am not satisfied with Codex in its current state. The limits are too restrictive, the usage calculation is not transparent enough, and the “hours” presentation creates expectations that do not match the actual experience.Please clarify:
- How exactly Codex usage is calculated.
- Why the limit is presented as “hours” if actual usable time can be much shorter.
- What specifically consumes the allowance so quickly.
- Whether OpenAI plans to make Codex limits more transparent.
- Whether paid users will get a lower-model fallback after the top-model allowance is exhausted.
- Whether users affected by this misleading presentation are eligible for credit or refund.
Codex could be a strong product, but the current limit presentation and real usage experience feel unacceptable for the price.
P.S. Another serious issue is that Codex appears to consume usage even while reading project/system context at the start of a session: repository instructions, project files, startup context, tool context, and other required information before the user even receives useful work.
If this startup/context-reading process consumes the same allowance, then the system is even less fair to paying users. A user should not lose a large part of the limit just because Codex needs to load and understand the project before doing the actual task.
If context size is a major factor, OpenAI should make this transparent and fair. Either limit the amount of context a user can provide before it becomes wasteful, clearly warn the user that the current project/context will consume a large part of the allowance, or exclude mandatory startup/project-reading overhead from the paid usage limit.
Right now it feels like Codex charges the user for everything: reading files, loading context, understanding instructions, choosing tools, and only then maybe doing the actual work. At this rate, it feels like even small interactions could eventually be counted against the limit. This is extremely frustrating and makes the product feel unpredictable and unfair.
The user needs a clear usage breakdown: how much was spent on reading context, how much on actual coding, how much on tool execution, and how much on model reasoning. Without that transparency, the “hours” limit is not meaningful.
150 messages “I would write 150 words” 30 messages, and the tokens run out sooner. Thank you for your reply, I realized your forum is interesting. I’ve drawn my own conclusion. Good luck in your difficult work.
Although, to be fair, it is not a bad idea — for $100, I might actually try it 
