After about a week of testing, I think I finally understand how the new Codex limit system works in practice.
I’m writing this post because many people are still trying to interpret the new system using the old mental model. They see that a single request consumed a surprisingly large percentage of the 5-hour limit and immediately assume something is broken.
But in practice, the new system makes much more sense if you stop thinking in terms of message count and start thinking in terms of reasoning time.
That is the key idea.
The core principle
Under the new system, the most useful practical way to understand Codex usage is this:
the longer the agent spends reasoning, the more of your 5-hour limit it consumes.
So the real question is no longer:
“How many messages do I get?”
The real question is:
“How many minutes of reasoning are included in my plan, and how much does each minute cost as a percentage of the 5-hour limit?”
Once you start looking at it that way, the behavior becomes much easier to understand.
1. A practical formula
A simple way to estimate usage is:
Cost of a request = reasoning time × percentage cost per minute
This is not an official formula, but from a practical user perspective it is the most useful way to estimate real-world Codex usage.
2. Plus plan on GPT-5.4
Based on testing, the Plus plan on the latest GPT-5.4 model appears to provide roughly:
about 40 minutes of reasoning per 5-hour limit window
That means, approximately:
-
40 minutes = 100%
-
20 minutes = 50%
-
1 minute = 2.5%
So on Plus + GPT-5.4, a good working estimate is:
1 minute of reasoning costs about 2.5% of the 5-hour limit.
3. Plus plan on GPT-5.3
In practice, GPT-5.3 appears to be more efficient than GPT-5.4.
Based on testing, a reasonable estimate is:
about 60 minutes of reasoning per 5-hour limit window
That gives us:
-
60 minutes = 100%
-
30 minutes = 50%
-
1 minute = about 1.66%
So on Plus + GPT-5.3, a practical estimate is:
1 minute of reasoning costs about 1.66% of the 5-hour limit.
This is why model choice matters so much. Even if the task feels similar, the effective drain rate can differ noticeably depending on which model you use.
4. Business plan
Now to the part that caused the most confusion.
After revisiting my earlier calculations, I think I now understand much better what went wrong in how many of us initially interpreted the Business plan.
The main issue is this: OpenAI appears to have described the difference between Business and Plus in a way that sounded much softer than the real user experience.
Many users understood the claimed “around 60% difference” to mean something like this:
Business should drain the limit about 60% faster than Plus.
That is the interpretation most people would naturally make, because users do not think in terms of abstract allowance math. They think in terms of what they actually feel during work:
how fast the percentage disappears.
But based on testing, that does not seem to be what OpenAI meant.
What they most likely meant was something different:
Business includes less total reasoning allowance than Plus.
That is a very different kind of statement.
To explain it simply, imagine Plus gives you 100 units of total allowance. If Business gives 60% less, that does not mean the limit drains 60% faster. It means Business gives you only 40 units instead of 100.
So the claim is about the size of the total budget, not the speed at which that budget burns.
And that distinction matters a lot.
Because once the total budget becomes much smaller, the exact same task starts consuming a much larger percentage of that budget. That is why users may hear “60% less” and expect a moderate downgrade, while in practice the plan feels dramatically worse.
Now here is the practical version using my current estimates.
If Plus gives about:
40 minutes of GPT-5.4 reasoning per 5-hour window
and Business gives only about:
12.5 minutes
then the problem becomes much easier to see.
This means Business is not just “a little smaller.” It means the total available reasoning time is far smaller.
Put differently:
-
Plus gives about 40 minutes
-
Business gives about 12.5 minutes
-
so Business gives only a small fraction of the working time that Plus gives
In other words, OpenAI’s framing appears to have been about the remaining total allowance, while users were thinking about how fast the limit disappears during real work.
Those are not the same thing.
And the gap becomes even more obvious when you look at the practical drain rate.
On Plus + GPT-5.4:
- 1 minute ≈ 2.5%
On Business + GPT-5.4:
- 1 minute ≈ 8%
So in real use, Business does not just feel “around 60% worse.”
It feels dramatically more restrictive, because the limit drains:
8 / 2.5 = 3.2× faster
That means the real-world drain is about:
220% faster per minute than Plus
And that is the part many users were never clearly told.
There is also another important point here.
The issue is not only that the difference appears to have been described using a softer metric. The issue is that even within that softer framing, the real reduction now looks worse than “around 60%.”
If my current estimates are correct, then going from 40 minutes on Plus to 12.5 minutes on Business means the actual drop is closer to:
about 68.75% less included reasoning time than Plus
So even the softer framing may have understated the real reduction.
That is why so many Business users felt that something was broken.
It may not have been a bug at all. It may simply have been a much more severe regression in practical usability than the wording suggested.
This also helps explain why Business users kept reporting that the plan had become nearly unusable, while those complaints were easy to dismiss if someone only looked at the softer allowance framing instead of the real per-minute drain.
So if my current estimates are correct, then on Business + GPT-5.4 the practical model is:
-
12.5 minutes = 100%
-
1 minute = about 8%
That means:
-
2 minutes = about 16%
-
4 minutes = about 32%
And once you look at it this way, the reports from Business users suddenly stop looking irrational.
They start looking completely predictable.
That is also why I think OpenAI needs to answer this more directly.
If the difference between Business and Plus was communicated in terms of included allowance, why was it not made equally clear that, in real usage, this would translate into a dramatically faster per-minute drain?
Because from a mathematical point of view, that framing may be defensible.
But from a user point of view, it was not transparent enough — and it made the regression feel far less serious on paper than it turned out to be in practice.
5. Business plan on GPT-5.3
If you switch to GPT-5.3, which appears to be more efficient, then Business gives roughly:
about 18.75 minutes of reasoning per 5-hour limit window
That means:
-
18.75 minutes = 100%
-
1 minute = about 5.33%
So on Business + GPT-5.3, a rough practical estimate is:
1 minute of reasoning costs about 5.33% of the 5-hour limit.
This also explains why the same task can feel noticeably cheaper on GPT-5.3 than on GPT-5.4, even on the same Business plan.
For example, a 4-minute run would roughly cost:
4 × 5.33% = about 21.3%
That is still expensive, but clearly better than GPT-5.4 on Business.
6. Important clarification: reasoning level itself is not the direct price driver
This part is very important because it is easy to misunderstand.
The reasoning level itself — low, medium, high, or very high — does not directly change the price of a run.
A 4-minute run costs roughly the same percentage whether the reasoning level was low or very high.
What actually changes is that higher reasoning levels often cause the model to think for longer.
So the issue is not:
“high reasoning is priced higher.”
The issue is:
higher reasoning usually leads to longer reasoning time, and longer reasoning time costs more.
That distinction matters a lot.
Many users understandably come away with the impression that higher reasoning is more expensive, when in practice the real cost increase usually comes from the extra time spent reasoning.
7. How to use this model in practice
This is the most useful takeaway from all of this.
If you want to manage your limits more intelligently, stop thinking in terms of message count and start thinking in terms of included reasoning minutes.
That gives you a much clearer planning model.
For example:
-
if you need the strongest model, use GPT-5.4, but expect it to burn through limits faster;
-
if you want more total working time, GPT-5.3 is more economical;
-
if a task is likely to take several minutes of agent reasoning, you can estimate the cost before running it;
-
if you are on Business, you should expect the limit to disappear much faster than on Plus.
Once you think in minutes instead of messages, the new system becomes much more predictable.
8. Why users were confused
I do not think users were wrong for assuming something was broken.
The real issue is that the new system was not transparent enough, so people had no clear mental model for interpreting the new behavior.
Now we do.
The most practical way to understand the current system is:
-
your real cost is tied to reasoning time;
-
each model has a different effective cost per minute;
-
Business includes much less than Plus;
-
and once you understand the approximate percentage-per-minute rate, the system becomes far more predictable.
9. Where Pro 5X fits into this
There is also an important clarification to make about the new Pro 5X plan, because it helps explain the current Codex limit system even more clearly.
What matters here is not just that Pro 5X is a paid upgrade. What matters is that, in practice, it gives a much larger amount of available reasoning time relative to Plus — and under the current promotion, that increase becomes large enough that limits can almost disappear from normal workflows.
Using Plus as the baseline:
on GPT-5.4, a 5-hour limit window appears to include roughly 40 minutes of reasoning time.
That means:
-
20 minutes = about 50% of the limit
-
40 minutes = about 100% of the limit
Now, the new Pro 5X subscription costs $100 and is designed around a 5× Codex allowance relative to Plus. But under the temporary promotion running until May 31, 2026, it effectively provides 10× the Plus allowance instead of 5×.
So in practical GPT-5.4 terms, that means:
-
Plus ≈ 40 minutes of reasoning per 5-hour window
-
Pro 5X during the promotion ≈ 400 minutes of reasoning per 5-hour window
That is roughly:
6.7 hours of reasoning time inside a single 5-hour limit window
From a practical point of view, that is enormous.
For many real work scenarios, this can already be treated as effectively removing the limit problem altogether — especially for users working in Fast mode.
This is also why Pro 5X is such an interesting option: compared to the older and more expensive Pro 20× for $200, the new $100 Pro 5X plan may actually feel more efficient or more attractive in practice, because under the current promotion it behaves like a 10× plan, while costing much less.
In other words, for $100, users can temporarily get a level of Codex access that is large enough to stop thinking about limits in everyday work and simply use the system at full speed.
That is why Pro 5X should not be viewed as just “another subscription tier.”
Under the current promo, it is much closer to:
a practical way to get rid of limit anxiety and work at maximum pace without constantly watching your percentage.
So if Plus is the baseline for understanding reasoning cost, and Business is the plan where the drain feels noticeably aggressive, then Pro 5X — especially during the 10× promotion — is the tier where limits begin to feel almost irrelevant for serious daily use.
10. Final conclusion
So the practical conclusion is simple:
-
Plus gives a useful baseline for estimating reasoning cost in terms of time;
-
Business provides a much smaller allowance, which is why the drain feels aggressive;
-
Pro provides such a large allowance relative to Plus that, for many users, it may feel close to unrestricted.
If your goal is simply to understand the formula, then think in reasoning minutes.
If your goal is to stop constantly worrying about limits, then Pro appears to be the most comfortable option.
That is the whole point of this post.
Important note
Everything above is based on practical testing and approximation, not on official OpenAI documentation. The numbers should be treated as a working user model, not as exact published plan specifications.
But as a practical framework for understanding how the system behaves after the April 9 change, this model seems to explain user reports surprisingly well.
