Codex Rate Limits Discussion Thread

Codex
787

Same issue here with GPT-6 Astra.

With GPT-5.6 Sol, my 5-hour usage window normally allows me to work for a little over one hour of continuous, intensive coding before I run into the limit.

With GPT-6 Astra, I’m seeing the entire 5-hour window consumed in approximately 4 minutes of actual work.

This is an enormous difference in usage consumption. Based on my experience, Astra is burning through the available quota more than 15x faster in terms of actual working time than Sol.

I’m not talking about a slightly higher consumption rate — the difference is so large that Astra becomes practically unusable for sustained development work, even though I really like the model’s capabilities.

It would be helpful to know whether this is expected behavior or whether there could be an issue with how Astra usage is being calculated against the 5-hour window.

Is anyone else seeing consumption this extreme?

I can also provide /usage screenshots or additional details if that would help investigate it.

788

I want to report that in just 4 minutes and 21 seconds, a task where I asked Astra to record a number of observations and make corrections to the code of an application currently under development was interrupted because my 5-hour limit had already been exhausted. Fast mode was not enabled, and I was using Medium reasoning. That seems like an excessive amount of quota consumed for such a short amount of time, especially without any conclusive results yet. I should also point out that I still had 77% of my 5-hour allowance remaining when I sent the prompt.

In addition, when I restarted with Astra on Low reasoning, it completed the requested actions in 15 minutes and 37 seconds. I have not yet been able to verify the result, but it was faster than Sol would likely have been. However, it also completely exhausted a fresh 5-hour window that had started at 100% quota.
Overall, Sol at High reasoning would probably have been slower, but I do not think it would have consumed nearly as much. Unfortunately, I do not have the means to properly compare the quality of the two results.
So, based on my experience with Astra, you really need a substantial usage budget and to know carefully when it is worth using it instead of less expensive models. I hope OpenAI’s teams are able to reduce Astra’s usage consumption, because in its current state, unless you have a very large allowance and a Pro plan, it seems difficult to use in practice.

789

I like this a lot as current setup forces weird timings for the cutoff.

790

I’ve been seeing the same issue and initially focused on how quickly Astra was consuming the available budget.

After looking at my own workflow, though, I realized that part of the problem was resource allocation. I was effectively using the strongest reasoning model for repository exploration, routine implementation, tests and other tasks that could be handled elsewhere.

I’m now experimenting with routing different parts of an engineering task to different agents/models:

Router → Context Scout → Premium Reasoning → Implementer → Reviewer

The goal is to minimize the repository context reaching the premium model and reserve it for decisions that actually require deep reasoning.

It doesn’t change Astra’s underlying usage limits, but it may make the available reasoning budget much more useful.

I’ve published the experiment as an open-source GitHub template called Token Gourmand under my GitHub profile robert-lukowski.

It’s still early, and I’d be especially interested in comparing real measurements from different workflows rather than making assumptions about the savings.

791

There are several related rate-limit issues being reported in this thread.

One is Astra consuming the Plus 5-hour allowance extremely quickly, which looks like a serious issue for Plus subscribers in its own right.

Another — and the critical one for me — is an apparent regression in GPT-5.6 Sol Medium usage efficiency (and possibly other models I haven’t used). Sol Medium had been my reliable baseline for substantial development work, but it now appears to consume the allowance dramatically faster than it did previously for comparable tasks.

I reverted to Sol Medium after the first abnormal Astra run specifically because it has historically been the dependable option, and therefore provided an obvious short-term workaround to Astra. The fact that Sol Medium is now also shredding the allowance is what has, effectively overnight, made Codex unusable for my workflow.

792

Dear Codex Development Team,
I am writing to share some feedback regarding the current Codex usage experience (Plus) and to ask whether the usage-limit system could be reviewed.

Since the release of GPT-6, I have noticed that my five-hour Codex usage allowance can be consumed extremely quickly. For the same or very similar tasks, the previous GPT-5.6 model appeared to use roughly half as much of the available quota, sometimes even less. I do not know whether this behaviour affects every user, but I believe many users have noticed how rapidly the available allowance now disappears.

I am not in a position to judge whether the increased runtime consumption is technically necessary. However, consuming an entire five-hour allowance in approximately 14 minutes does not seem like a normal or practical user experience (only one task). In situations like this, the usage limit significantly restricts Codex and prevents users from making effective use of it.

I have also noticed a separate performance issue on the web version: when a conversation becomes very long, Safari can become extremely slow and unresponsive.

I hope the development team can review these issues and consider adjustments that would improve the overall user experience. I would also welcome comments from other members of the community who are experiencing similar problems.

Thank you for your time and consideration.

793

I share the same frustration regarding the consumption of usage limits. What is particularly aggravating is that my quota gets used up just trying to get the AI ​​to correct its own mistakes.

For instance, when I asked it to create an Excel document following a specific GUI layout, it arbitrarily changed the order of elements. When I asked it to fix the order, it proceeded to alter the items themselves. Despite this, it would report the task as “verified” and “complete,” meaning my usage quota dwindled while I was stuck in a loop of pointing out errors and requesting revisions.

I pay for this service to make my work more efficient, yet in reality, I end up with more work: spotting the AI’s errors, explaining them, and overseeing the corrections. It is a rather ironic form of “evolution” that hitting the usage limit is a more certain outcome than actually getting a finished deliverable.

I would like the new development team to verify not just the rate at which the quota is consumed, but also whether the requested work is actually being completed successfully within those limits. What users want isn’t apologies or repeated reports of completion—we want usable deliverables.

794

I am having the same issue. I have went through 70% of my weekly limit on the Pro max tier with only about 5 hours of runtime on a query that should not consume so much credits.

795

I think codex should have soft limits instead of hard limits so the ai stops while finishing in a good posiotion to contue rather then a full hard stop

796

Hi,

I’m experiencing what appears to be unusually high weekly credit consumption on Codex and wanted to check whether there may be an issue with how usage is being calculated on my account.

I’m currently subscribed to the $200/month ChatGPT Pro plan, so I’m using the highest individual subscription tier available to me.

I believe my weekly Codex usage reset approximately yesterday, although I don’t want to state the exact reset date with certainty in case I’m remembering that incorrectly. What I do explicitly remember is checking my usage yesterday and seeing that I still had more than 90% of my weekly credits remaining.

I have now had approximately 9 hours of Codex runtime, and my weekly credit allowance is down to only 14% remaining.

That means roughly 75%+ of my entire weekly allowance has been consumed over this period. Even allowing for some uncertainty around the exact reset time, the rate of consumption appears extremely high.

I have primarily been using Astra 6 with Extra High reasoning, so I understand that this configuration should consume substantially more credits than lighter models or lower reasoning settings. However, consuming the large majority of a weekly Pro allowance after approximately 9 hours of runtime seems disproportionate, particularly on a $200/month subscription.

I use Codex for substantial software-development tasks, including longer autonomous coding and research sessions, so I fully expect meaningful usage. I’m not reporting this simply because credits are being consumed. My concern is the magnitude and speed of the consumption relative to the weekly allowance available on the Pro plan.

Could someone from the OpenAI team please check:

  • Whether my weekly credit usage is being calculated correctly

  • Whether Astra 6 Extra High currently has an unusually high or recently changed usage multiplier

  • Whether there has been any recent change to how Codex usage is calculated

  • Whether autonomous sessions, tool calls, retries, context processing, or other session operations could be causing substantially higher consumption than expected

  • Whether there is a way to see a detailed breakdown of what specifically consumed the weekly allowance

  • Whether approximately 75%+ of a Pro weekly allowance being consumed after around 9 hours of runtime is considered expected behavior

Again, while I may be slightly off about the exact weekly reset date, I am confident that when I checked my account yesterday I was still in the 90%+ range for weekly credits remaining. I am now at 14% remaining after approximately 9 hours of runtime.

If this is expected behavior rather than a bug, I would appreciate clarification on approximately how much Astra 6 Extra High usage the Pro weekly allowance is intended to support. At the current rate, a $200/month Pro account could exhaust essentially its entire weekly Codex allowance in around a single day of substantial development work.

I’m happy to provide screenshots, session information, or any other account-specific details privately if they would help investigate this.

Thank you.

797

you turned on Extra High on Astra probably. Are you trying to disprove a theorem or what?

798

What I’m seeing more often than not, is that these $20/$100/$200 “Pro” subscriptions are really just Demo versions of a tool that could do the job it says it could do. Those 5x/20x rates mean nothing as so many of us have realized. For serious work, complex tasks, full-on projects especially on a short deadline, is nowhere near enough. So you need to switch to credit-based purchases, through which you’ll soon realize you’ll be heading to 4- even 5-figure costs.

So yeah, nothing’s broken, the technology is quite capable, but unfortunately it is very expensive at this time.

799

I tried using terra it gives me way more usage while barly losing quality of work if you don’t need that much intelgence terra is a good middle ground

801

My Codex usage suddenly dropped from 78% to 3% today, September 9, 2026.

At around 7 PM, I still had roughly 78% of my weekly usage remaining. I’ve only been running a single GPT-6 Astra thread with one agent doing some minor work. Then, out of nowhere, I got a popup saying I only had 3% of my weekly usage left.

That doesn’t seem possible at all. There’s no way a single thread doing light work could have used around 75% of my weekly allowance that quickly. This looks like a usage tracking or billing issue.

I opened a support ticket through the support, page but this has now automatically been closed!

Screenshot 2026-09-09 at 7.26.21 PM
Screenshot 2026-09-09 at 7.26.21 PM1452×384 66.6 KB

802

I am 20x user and i noticed this within few hours I am out of credits

image
image816×346 11.5 KB

803

My weekly quota was about to be exhausted, close to zero, and then shot to almost full, but not quite! Something fishy is happening.

804

Now my quota is back to 6%! What is happening?

805

Hi all,

This might be related to the incident OpenAI is investigating about unexpected Codex usage limit resets. It hasn’t been confirmed whether it includes the sudden drop you’re seeing, but you can follow updates here:

Investigating unexpected usage limit resets

806

I lost the 26% I had left, and now it looks like it’s back,. BUT they moved up the reset day!

807

Environment

Surface: Codex CLI

Endpoint or feature: Usage quota reporting via the /status command

Model: [Enter the model shown in your Codex session]

SDK and version, if applicable: N/A
Codex CLI version: [Run codex --version and enter the result]

Bug

What happened, and what did you expect?

My usage quota was reset today. Shortly before midnight, the /status command showed that I had more than 90% of my quota remaining. Shortly after midnight, the remaining quota
unexpectedly dropped to approximately 2%, even though I had not performed enough work to account for such a large amount of usage.

I expected the remaining quota to stay close to the previously reported value, minus only the actual usage incurred after the reset. Please investigate whether the reset was applied
incorrectly or whether the quota/status calculation is inaccurate.

Reproduction

Minimal code, request, or steps to reproduce:

  1. Open a Codex CLI session.
  2. Run /status shortly before 00:00.
  3. Observe that the remaining quota is above 90%.
  4. Wait until shortly after 00:00.
  5. Run /status again.
  6. Observe that the remaining quota has dropped to approximately 2%, without usage that could reasonably explain the decrease.

Diagnostics

Exact error and HTTP status:

No HTTP error was displayed. The issue is an unexpected usage-quota decrease reported by /status.

Request ID, if available:

Not available.

Add any request ID from the affected session if you can find one.

When it happened (with timezone) and how often:

It happened around midnight between September 9 and September 10, 2026, in the Asia/Bangkok timezone (UTC+7). Before 00:00, /status showed more than 90% remaining; shortly after
00:00, it showed approximately 2% remaining.

I have observed this issue once so far.

image
image953×481 67 KB

image
image953×481 63 KB