Severe model degradation and persistent “Selected model is at capacity” errors

ChatGPT Bugs
, ,
1

[Account Issue] Pro 20x account became unusable after upgrade — severe model degradation and persistent “Selected model is at capacity” errors

Hello OpenAI team and community members,

I would like to report an issue with my ChatGPT Pro 20x account that has prevented me from using the service normally.

This is a personal account registered with a Google email that I have used for several years. My subscription history is: ChatGPT Plus → 5x → Pro 20x. My main usage is AI development work, especially AI video prompt generation workflows and local Codex-assisted development. I also use ChatGPT Web / Work mode to handle other tasks in parallel.

Before this issue happened, my account usage was relatively heavy. My previous 5x Pro allowance was usually consumed within around two days. I do use proxy nodes, but before the problem started, my IP environment was relatively stable. Most network changes were only performed after the issue appeared as part of troubleshooting. This account is for personal use only.

Timeline (US Eastern Time, EDT)

September 10, 2026 — 07:00 AM EDT

I upgraded my account directly from the official ChatGPT website to Pro 20x.

The payment was completed using Philippine currency.

After the upgrade, the account worked normally.

September 10, 2026 — 08:30 PM EDT

I started testing the upgraded account.

I tested different models for AI image/video-related tasks and development assistance.

At that time, most models performed normally. Sol showed some noticeable quality degradation, but the other models were generally working as expected.

Overall, the account was usable.

September 12, 2026 — Around 02:00 AM EDT

The issue started around the scheduled reset period.

Before this point, my conversations had already started showing some instability.

After this point, the model behavior changed significantly.

The model could still understand my questions, but it often failed to continue with the actual requested work.

The typical pattern was:

  • It understood the problem.
  • It explained what the problem might be.
  • It repeatedly said things similar to “I already understand where the problem is.”
  • However, it did not continue with the actual execution.

The experience was very different from the previous model behavior and felt similar to using a much weaker model.

Because my main workflow depends on AI-assisted development, this directly affected my ability to work.

Testing Different ChatGPT Interfaces

After noticing the problem, I tested different OpenAI products to determine whether this was a general issue or an account-specific issue.

ChatGPT Normal Chat

The normal ChatGPT interface was still usable.

The model capability itself appeared mostly normal, but the response speed became significantly slower.

ChatGPT Work Mode

Work mode was still accessible, but the model behavior appeared significantly degraded.

I used a third-party model tracing tool as an additional reference during troubleshooting. The result suggested that successful requests might have been routed to a lower-capability model.

I understand that third-party tracing is not official evidence. I am only mentioning this as an observation and hope OpenAI can verify the actual model routing and account status internally.

September 12 - September 14, 2026

After the issue appeared, I started troubleshooting:

  • I stopped active usage for a period of time.
  • I logged out Codex from all devices.
  • I stopped parallel Codex sessions.
  • I submitted reports through OpenAI Support.
  • I contacted support through the Help Center.

September 14, 2026

I received a response from OpenAI.

The response mentioned that an account warning had been removed.

However, I was not sure what this meant.

I could not determine whether:

  • the account had been fully restored;

or:

  • only the warning status had changed while other restrictions remained.

The response did not directly explain the model capability degradation issue or why my account behavior was different from other accounts.

The human support responses also appeared to provide general troubleshooting information rather than addressing my specific situation.

Codex Situation

On September 14, Codex showed a further change.

From:

September 13, 2026 — 08:00 PM EDT

to

September 14, 2026 — 02:00 AM EDT

Codex using 5.6 Sol was still able to work.

However, the capability was significantly reduced. It could continue development tasks, but the quality was similar to a much weaker model.

After:

September 14, 2026 — Around 02:00 AM EDT

Codex started occasionally returning:



Selected model is at capacity.
Please try a different model.

After:

September 14, 2026 — Around 04:00 AM EDT

Codex became almost completely unusable.

Almost every request returned:



Selected model is at capacity.
Please try a different model.

I submitted another support request regarding:

  • abnormal model capability;
  • inability to use Codex normally.

I also tested with a cleaner residential IP environment.

The issue remained.

September 15, 2026

I received another email response from OpenAI.

The response stated that the previous decision would remain unchanged.

However, I am unclear whether this means:

  • the warning removal remains unchanged;

or:

  • the current account state remains unchanged.

The response still did not explain:

  • why my account behavior changed;
  • why Codex became unavailable;
  • whether my account currently has any abnormal status.

Account Comparison Test

I tested another Plus account under the same environment:

  • same computer;
  • same network;
  • same client environment.

Result:

My Pro 20x account

  • Codex frequently unavailable.
  • Persistent capacity errors.
  • Significant model quality degradation.

Another Plus account

  • Normal model access.
  • No similar issue.

The other account’s development token consumption is slightly lower than mine.

This comparison makes me believe the problem may be related to my account state rather than my local environment.

Current Situation

As of now:

  • OpenAI Status page shows that the latest ChatGPT service issues have been resolved.
  • However, my account continues to experience problems.
  • My ChatGPT Work mode also entered a state where it frequently returns capacity errors and has become unusable.

Since the latest reset:

  • I have only used around 7% of my Pro 20x allowance.
  • During this period, I only had around one and a half days where the model worked normally.

I upgraded to Pro 20x because I needed a more reliable development experience.

However, currently I feel that I paid for the Pro 20x service but cannot consistently use the service normally.

Possible Reasons I Considered (Not Confirmed)

I am not claiming any of these are the actual cause. I am only listing possible factors that I hope OpenAI can verify.

1. IP changes

One possibility is that automated security systems may have affected my account.

However:

  • My IP environment was stable before the issue.
  • Most IP changes happened only after the problem started, when I was troubleshooting.

2. Multiple Codex sessions

After upgrading to Pro 20x, I tested running Codex on two devices under the same network.

The purpose was:

  • one device for the main project;
  • one device for a related sub-feature.

The number of active sessions was usually around 2-3.

After the issue appeared:

  • I logged out all Codex sessions;
  • stopped parallel usage;
  • left the account unused for a period of time.

The problem did not recover.

3. Subscription tier resource allocation

I noticed some Pro users reporting similar model degradation and capacity problems.

I understand that large-scale AI services can experience high demand.

However, I would appreciate clarification on whether different subscription tiers have different resource allocation behavior during peak periods.

I do not have evidence to confirm this, but I hope OpenAI can clarify.

My Request

My request is simple:

I hope OpenAI can help investigate and restore my account to normal usage.

I would appreciate help checking:

  1. Whether my Pro 20x subscription and model access are working correctly.
  2. Whether there is any abnormal account state, incorrect configuration, or restriction affecting my usage.
  3. Whether my Codex access is correctly configured.
  4. Whether my account can be restored to the normal experience expected from a Pro 20x user.

I am not asking for special treatment.

I only hope to understand why my account cannot currently work normally and receive a clear solution.

If there is a technical issue, I hope OpenAI can acknowledge it and provide a recovery path.

If other users are experiencing similar issues with:

  • model capability degradation;
  • Codex capacity errors;
  • Pro plan account problems;

please share your experience so we can better understand the scope of this issue.

Thank you.

2

This subcategory has specific requirements for posting:

I am moving this topic to a different category.

3

I’ve encountered the same issue; it has been going on for a week, and despite reporting it twice, it still hasn’t been resolved.

4

As of now, I’m still waiting for a response from OpenAI Support.

5

image
image1178×578 18 KB

I just renewed my 20x subscription two days ago, but it’s been unusable for the past few days. They say the problem has been fixed, but I don’t know exactly what’s been fixed. This is very frustrating.

6

I have the same exact issue…

image
image780×200 10.2 KB

7

I need it to complete my endless work, but I’m having the same issue too.

8

I’m a Pro 20x user as well, and this week has been the worst experience that I’ve had with Codex since starting. I am buying the Claude Code $100/month subscription right now because the experience this week has been terrible.

I’m not sure of the final cost yet, but I’ve spent about $500 in credits this week because my usage limit was exhausted so quickly. Not only that, the actual performance of the models was so bad. I had to go back and steer them constantly. It’s just been a terrible week.

9

I will keep ongoing records of my account operations and official support responses to track progress.

On September 15, I submitted an appeal to human support through the Help Center and have not yet received a reply.

Today I ran tests on all models under the web Work mode. The first message each day responds with normal capability, but all subsequent conversations suffer degraded performance. Chat mode maintains normal capability. It has been 15 hours since my last login; I logged out all my active accounts 15 hours ago. I am continuing to log out to cool down the account.

From past experience, the Work models behave like a water tank, with OpenAI acting as a dripping faucet. If I leave it unused, a small amount of usable quota accumulates. With coding tasks, I can work for roughly 1–2 hours daily on the 5.6 Luna model. The 5.6 Sol model appears to be a facade. Since Astra launched, I have never been able to access a properly functioning 5.6 Sol. I have used third-party capability testing tools to verify routing.

My current plan is to cool down the account for 5 days and see if performance recovers. I will continue to follow up on this matter and watch for a clear official response, though I suspect their reply will most likely just ask me to review their user documentation and guidelines.

If you are facing the same issue and are a legitimate user rather than an abuser, please join us and speak up. We need to gain a fuller understanding of this matter and the official stance.

10

image
image976×773 21 KB

same issue

11

it’s a shadow ban on your account. for me and some people i know, it happens around 10pm to around 6am ET everyday with endless “model at capacity” warning, which is a blatant lie (it’s not really at capacity, just an artificial blocker to prevent you from using anything). my 20x account can only use chat mode during the affected hours.

from what i can see on X, or here, or elsewhere, the affected group is overwhelmingly chinese speaking user, so it could be a targeted discriminational purge. very gross practice on openai’s side.

12

You’re correct. Most of the information I’ve collected points to X20 problems hitting users in China, Japan, and the Philippines. If subscribers outside these regions are facing the same issue, feel free to drop a comment.

13

I would be very careful with the jump from the real symptom to “targeted discriminatory purge / shadow ban.” The evidence in this thread does not establish that.

There is an official datapoint worth separating from the speculation: OpenAI Support posted on Sep 14 that access to some models/features may be temporarily limited based on account activity even on a paid subscription, that they cannot provide additional details about those checks, and that access is automatically reassessed. That makes account-specific access treatment a documented possibility.

But that is still a long way from proving geography, language or ethnicity is the predicate.

There are also obvious counterexamples to an Asia-only theory. In the older capacity thread, a US Pro 20x user reported repeated Selected model is at capacity errors. And a user’s residential IP/country is not the same thing as the backend region that executed a particular turn; current long-running Chat/Pro controls I have inspected have landed in both US and European SAServer regions.

So the defensible claim right now is narrower: multiple paid accounts are reporting capacity/access/capability problems, and some reports look account-specific even under the same local environment. The hidden predicate is unresolved.

If people already have diagnostics, the useful comparison is dated plan/model/effort + exact error + native request/region metadata where available, ideally an A/B from the same machine/network. That can test an account-treatment hypothesis without turning nationality or language into the explanation before the evidence gets there.

15

I just signed up for another Plus account. It’s faster and produces higher-quality replies than my X5. I did the pelican test again, and the outcome was shockingly better. I’m pretty frustrated, honestly.

16

RIP our account​:wink:
https://community.openai.com/t/issue-selected-model-is-at-capacity-please-try-a-different-model

17

image
image1149×930 62.7 KB

What’s wrong with my account?

18

Just got word that GPT is having another large-scale outage. Has it hit your account yet?

19

I was getting connection interruptions in a project chst I was using 5.6 thinking on extra high and it went through thinking and doing the work 35 minutes later it gives a connection interruption instead of finishing the task

20

As of now, human support still hasn’t replied to me. It seems my account’s risk control has been escalated. When I simply sent “Hello”, it tried to reconnect five times, then returned this error message:

stream disconnected before completion: An error occurred while processing your request. You can retry your request, or contact us through our help center at help.openai.com if the error persists. Please include the request ID xxxxxxxx