Codex Rate Limits Discussion Thread

Codex
252

The reason you only ran into this problem now, and not earlier when other people in this thread were already writing about it, is that the new Codex request pricing system only came into effect for the Plus, Pro, and Enterprise subscriptions on April 9.

That is likely why you only noticed such dramatic changes now. As I said earlier, OpenAI made not only the Business subscription worse, but also the Plus subscription — just not as severely, though the difference is still noticeable.

For example, the fact that you burned through 60% of your limit in 20–30 minutes is completely normal under the new pricing model they introduced. It is not an error and not a bug, and I already explained how it works in my post above.

You may not have read it carefully enough, so I would encourage you to go through it again. You may find some valuable answers there.

The post I wrote is not a text generated by artificial intelligence; it is the result of my own research. However, I do not speak English well enough to write it on my own, so I used AI to translate my Russian text into English and, in some places, possibly smooth out the phrasing so that it would read more naturally for an English-speaking audience.

1 Like
253

bro you’re trying to trouble shoot me with out even looking at my codebase or the command I sent last night.

I’ve used it every waking hour since the changes, and it started last night, which was well after the changes.

I’ll defer to my own investigation for the time being, thank you, though.

254

I don’t think the answer is actually that complicated, although this is just my personal view.

From a business perspective, I think the key difference is how people actually consume the included limits in ChatGPT versus Codex.

Most ChatGPT users probably consume only a small fraction of the usage that is technically included in their subscription. In other words, the allowance is there, but the average user never comes close to fully using it.

To make it simple, imagine that a ChatGPT Plus subscription includes 100% of some available usage budget, but the average ChatGPT user only consumes, say, 10% of that in practice. For the company, that is a very comfortable model, because most users are not pushing the system anywhere near the maximum.

Codex, in my view, is used very differently. People who use Codex are much more likely to push it hard and burn through a large share of the included allowance. So instead of an average user consuming a small fraction, you may be looking at something more like 50% on average, or even close to full exhaustion for heavy users. And that starts to hit OpenAI’s economics much harder.

That is why I do not think ChatGPT and Codex can be compared purely on the basis of “same infrastructure, same tokens.” From a business standpoint, what matters is not only the underlying compute, but also the average intensity of usage across the user base.

We recently saw something very similar with Anthropic. Once people started using their subscription through OpenClaw and effectively burning the full included limits through agent-style workflows, it began to hurt their business economics. And as you probably know, Anthropic eventually prohibited using subscriptions that way: if you connect a subscription through OpenClaw instead of using the API, the account can be banned.

So my view is this: the issue is probably not that ChatGPT is inherently cheap for OpenAI while Codex is inherently expensive. The issue is that Codex users, on average, consume their included limits far more aggressively than ordinary ChatGPT users do. And that is probably what forced OpenAI to change the economics of Codex inside these subscription tiers.

1 Like
255

The point is that I am confident in my calculations. This is no longer just a hypothesis that still needs confirmation; it is a clear working formula that lets you calculate the cost of a request against the 5-hour limit based on three things: the model you use, the subscription tier on your account, and the amount of reasoning time the agent spent on the task.

In other words, that is already enough to estimate what percentage of the 5-hour limit should have been consumed.

In fact, for the past few days, I have been doing this on purpose out of curiosity: before even checking how much limit I have left, I first calculate for myself how much should remain after the request. Then I compare it to the actual number. So far, in my experience, my calculations have matched every time.

That is why, when you said that you used 60% of the 5-hour limit in 20–30 minutes, it did not surprise me at all. Under the current system, that is exactly the kind of result that makes sense for a Plus subscription.

The important thing to understand is this: under the new system, the size of your repository, the size of the input, and the size of the output matter far less than people think.

What really matters now is only one thing: how long the agent spent reasoning.

That reasoning time is what determines how much of the 5-hour limit gets consumed.

As I already explained above, under the Plus subscription, if you use GPT-5.4, the 5-hour limit appears to include only about 40 minutes of agent reasoning. If you use GPT-5.3, it is about 60 minutes.

That is why a situation where 20–30 minutes of work consumed around 60% of the limit is completely consistent with this model.

And this is not just something I made up. It is also fully consistent with the official explanations of how the new system works.

So I do not need to look into your code in order to estimate how much of the 5-hour limit should have been consumed. In the new model, the decisive factor is not the complexity of the project itself, but the amount of time the agent spent reasoning.

256

It’s an unpleasant habit that, to introduce a more expensive plan (here the $100 plan), AI providers downgrade the baseline plan (here the $20 plan). This time the downgrade is really harsh; basically Plus has become 5 times more expensive overnight.

3 Likes
257

I wanted to share my unhappiness with this change. I feel like I fell for a bait-and-switch/rug-pull with buying credits. I recently bought $40 worth of credits, and this change consumed almost all of them in ~1/5th the time it took before the rate limit / pricing change. IMO - you effectively changed the value of something I already purchased, with no recourse for the customer. It would have been better if existing/previously purchased credits continued to follow the old model.

I intended to purchase a Pro plan because codex struck the right balance between price and capabilities, at least for my use-cases. After this change, the value of all the consumer-facing plans decreased significantly without any adjustment in price. Why would I upgrade now if I’m going to have to buy credits and burn through limits/quotas ~5-6x faster than just a couple weeks ago?

Like many others, I’m burning through my 5h limit in 20-40 minutes using GPT-5.4 high. I understand other models are cheaper, but I use this for most of my workflows due to the quality and task adherence. Now, exhausting the 5 hour quota/limit will use ~40-60% of my weekly limit. Why even have a 5 hour limit if you can only exhaust it twice a week?

From a user perspective, you’ve materially changed the experience and value proposition. I understand that OpenAI needs to stay competitive and may need to make adjustments to pricing models, but at this rate I feel like I’m being forced into an API / credit model while also paying a subscription.

I really hope the OpenAI team takes all of our feedback seriously and makes a meaningful adjustment.

2 Likes
258

Same thing just happened to me. I had ~60% usage left, asked a simple coding question in Codex, and within minutes it was completely gone. Definitely feels like something changed with how tokens are being counted. This doesn’t match the actual size/complexity of the request at all.

1 Like
259

I am very disappointed with the new Codex pricing and usage policy.

The change feels abrupt and much worse for real day-to-day work. Tasks that used to be completely reasonable under Plus now burn through limits incredibly fast. In my case, even a single request like adding unit tests can consume a huge portion of the 5-hour limit. That makes Codex much less practical and much less predictable than before.

What is most frustrating is not only the higher effective cost, but also the loss of confidence in using the tool for normal development work. When you start a coding task, you should not have to worry that one larger prompt will suddenly drain your available usage.

A lot of users are clearly unhappy, and I feel exactly the same. This change seriously reduces the value of the subscription for anyone who uses Codex for real software engineering tasks.

I really hope OpenAI reconsiders this policy, improves transparency around usage calculation, and brings back a pricing model that feels fair and usable.

1 Like
260

Dear @OpenAI_Support ,

We all really love you, and the products you make are great.
But what is going on with the silence for a week? You can’t just ignore the hottest topic in community now…give us peasants at least one-liner explanation…Sam, hit that “Reply“ button, stop refreshing the page every 5 minutes!
A bunch of angry users is no good for IPO! :smiley:

261

I just created a new account gonna test it with a single prompt….

262

I have created a new account to test this. First single coding task which took 77k token and it is already consumed 42% of my weekly usage limit, I am literally shocked :exploding_head:

I believe there is definitely something wrong with the usage calculations. cc: @OpenAI_Support

image
image2530×522 51.6 KB

263

Yeah, it turns out I just aimed too high and gave instructions that were too complex for the system…

Not the first time I’ve done that…

whooops.

^.^

264

Since OpenAI is promoting the $100 plan at x10 (for Codex), they can keep it at that level and recalibrate Plus at x2. This recalibration would make the $100 plan appealing while allowing plus users reasonable access. After all, Sam Altman publicly criticized Anthropic’s elitist pricing model, so I believe OpenAI is championing a more inclusive approach.

265

I’m really disappointed that OpenAI essentially pulled the rug from the plus plan. The 5h usage burns up so quickly that you can miss it if you blink. I’ve cancelled my subscription and I’m not going to sign up for a more expensive plan only to have that one rugpulled as well eventually.

2 Likes
266

(post deleted by author)

268

I already explained in my previous post how the new system work

1 Like