"Selected model is at capacity" is an error caused by congestion on OpenAI's side (server overload), not by a problem with your PC or your usage limits. Work through it in this order: wait a little and retry, check OpenAI's incident information, and if you need to hurry, try a different model.

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

Check route when work stops

01 Save the request and wait

Keep the prompt you entered and the last completion report. Don't press repeatedly; retry after a short pause.

02 Check incidents and remaining usage

Check OpenAI Status and your usage separately. Incidents can happen even when you have usage remaining.

03 If you're in a hurry, consider an alternative

Choose another available model yourself. If several models are affected, switching may not fix it.

Recommended order. For failures that occur across several models at once, switching to another model may still not stop the problem.

What's happening? What the official incident records show

The literal translation is "The selected model is at capacity. Please try a different model." "Capacity" here does not mean your PC's memory or disk space. OpenAI's official status page records service-side incidents that come with this message.

June 16, 2026: Codex capacity error

OpenAI Status lists the Desktop, Web, API, CLI, and VS Code extension as affected. It reports that a mitigation was applied and service was restored (official incident record).

July 9, 2026: same message across several models

The body of the OpenAI Status entry contains the same full error message shown at the top of this article. It states that the error occurred across multiple models, and service was later restored (official incident record).

These two cases show that "the same message can appear because of a service-side incident" and that "switching to another model may not improve things." (Dates are as shown on the official pages.) Neither record states the root cause, such as a shortage of GPUs.

How this differs from usage limits, 401, and thread not found

Work stopping may look similar, but you check different places. When you see a capacity error, first combine the full text on screen with your usage status. Several different problems can occur on the same account.

Message or situationWhat to checkNotes on judgment
Selected model is at capacityOfficial incident information, time of occurrence, and the model you selectedThe message alone does not mean you have run out of your own usage.
Usage limit or reset-waiting noticeRemaining balance and reset time on the usage screenCheck your remaining balance, then decide whether to wait or how to continue.
401 Unauthorized
Incorrect API key provided
Sign-in method and active accountThis is an authentication problem. It is handled differently from a capacity error.
thread not foundLoading and resume state of the target chatThis is an error saying the chat can't be found. Don't treat it as model congestion.

Check your remaining usage on the dedicated page

OpenAI's pricing and usage limits guide recommends checking your current limits on the usage dashboard. If Codex CLI is running, you can also use /status. If you check right after the error occurs, it is easier to compare your remaining balance at that moment with the error.

The official guide also explains that if you reach a usage limit while a task is in progress, that task can continue under fair-use constraints. So "there was an error in the middle, but a completion report appeared later" alone cannot tell you whether you reached a limit. Conversely, service-side incidents can occur even when you have plenty of usage left. For how plan allowances and additional credits work, see our ChatGPT Pro plan comparison.

401: check the authentication method first

The Codex authentication guide separates signing in with ChatGPT from signing in with an API key. In the desktop app, you can check the active account and API key status from the profile menu. In the CLI, use codex login status.

If you are signed in with ChatGPT but see "Incorrect API key," check which credential was rejected. As a fix for a capacity error, you do not need to sign out or delete authentication files. If you use an API key, standard API pricing applies separately from your ChatGPT subscription. Also, on September 25, 2026, a separate incident caused 401 and 502 errors in Codex when signed in with ChatGPT due to a false positive in an internal authentication check, and an official root-cause report was published (this is separate from the capacity error).

If thread not found appears, see Investigating and fixing "thread not found" in Codex.

Treat API 503 errors and app messages separately

The OpenAI API error codes guide describes HTTP 503, service_unavailable_error, and server_is_overloaded, which indicate temporary model overload. A 429 relates to request rate or usage limits, and a 401 indicates an authentication error.

Don't determine the HTTP code from the app's wording
The API documentation is a useful reference for distinguishing errors, but Codex can show the same message through paths other than 503. The phrase "at capacity" on screen alone does not establish a 503.

Steps to get your work back on track

The following steps are based on OpenAI's official incident records, usage limit guide, and troubleshooting documentation. Because the cause is congestion on the server side, the basic approach is to save your work and wait.

1. Save your request and the last completed work

If the prompt you sent is still in the input box, copy it. Save the last completion report and the state of any changed files. Before closing or recreating the chat after an error, note what you asked for and how far it got. This makes resuming much easier.

An error message alone does not tell you whether file edits or commands were never run at all. If you develop with Git, check the diff or run git status to see the current state, and then request the same change again. For work that involves publishing, sending, purchasing, or similar actions, don't repeat the same instruction without checking the result first.

2. Wait a while and check the official status

Open OpenAI Status and check for incidents affecting Codex or model selection. Look not only at the current display but also at whether the time of the error falls within an incident's time window.

For API overload, the official guidance is to wait at least as long as the Retry-After value if one is provided, and otherwise to increase the interval between retries. The app does not show a wait time, so wait a little before retrying rather than sending requests repeatedly at short intervals. While an incident is listed, refer to the official recovery information.

The status page is aggregated information for the whole service. Even if no incident is listed, you don't need to assume that your PC is broken.

3. Check usage, and consider another model if you need to hurry

Check your remaining balance and reset time. If reaching a limit is explicitly indicated, follow that notice. When you see only the capacity error, don't treat buying additional credits or paid resets as a fix. Restoring your own usage allowance and whether the model can accept requests are separate issues.

Prioritize quality and continuity

If you want to resume with the same model, wait for service to recover. Use the waiting time to organize requirements, diffs, and checks that haven't been done yet.

Prioritize moving forward quickly

Choose another model that is available and resume with a small task. If the failure spans several models, switching may still not stop it.

The official model selection guide explains that in the desktop app you can change the model using the model and reasoning effort controls below the input box. In the interactive CLI, use /model. Available models vary by account and client.

Choosing a different model also changes the nature of the answers and the usage. Switching to the top-tier model does not necessarily avoid capacity errors. When you take over complex implementation or design work, check the output after the change with diffs and tests. For guidance on choosing, see also GPT Sol generation comparison and how to choose.

4. Separate the state only if the app itself is frozen

Distinguish between getting a capacity error and the app not responding to input, the device, or the screen. The official troubleshooting guide explains how to handle chats that seem stuck: check for pending approval prompts, check the terminal with basic commands, and start a new chat with a small request.

If the terminal stays frozen, the guidance is to wait for the chat in progress to finish and then restart the app. This applies when the app itself is frozen. Server-side congestion is not resolved by restarting. If other chats are running, check their state first.

If you resume in a new chat, briefly hand over the goal, the working folder, the changes already made, and the unfinished items. Don't send just "continue," assuming the new chat knows everything from the old conversation.

Information to keep if it happens again

If the error recurs, keep evidence in a form that lets you investigate the cause later. Recording the conditions before you change settings several times makes it easier to see under which conditions it failed.

Checklist for support requests and recurrence records

  • Date and time of occurrence, with time zone
  • Where you used it (app, CLI, IDE, etc.) and its version
  • The model you selected, the reasoning effort, and the speed setting
  • The full error text, and the actions taken just before it
  • Remaining balance and reset time at that moment, and official incident information
  • Whether the error still appeared on another model after waiting and retrying

Even if things were fixed right after you changed the model, the service may simply have recovered at the same time. In your records, don't write "it worked after I switched." Write the time of the change and the result as they happened.

The official guide explains how to send feedback by typing / in the message input. When you send it from an existing chat, you can choose whether to share the conversation. Remove API keys, email addresses, private conversations, and internal information from any logs or screenshots you send. You don't need to publish key values or the authentication files themselves.

Reports are easier to investigate when they say not only "it keeps breaking," but "at this date and time, with this model, this message appeared, and the remaining balance displayed was this." If you obtained a communication code or request ID, that is useful material to pass to support.

Our own logs also classified it as server overload

On October 1, 2026, we checked the Codex execution history in our environment (the Windows app), where this error had recurred repeatedly. The 12 occurrences across five chats between 13:10 and 15:24 UTC on September 30, 2026 all had the same full text and were accompanied by codexErrorInfo: serverOverloaded. They were recorded as server overload, not as usage limits being reached. The records did not include the reason for the overload or an HTTP code.

The error definitions in Codex CLI 0.159.2, bundled with the app, also map this full text to ServerOverloaded, a classification separate from usage limits. The stream receiving process converts server_is_overloaded into an overload error, and the display conversion process links it to the full message shown at the beginning.

Summary: check the error type before acting

If "Selected model is at capacity" appears, save your request, wait a little, and then check the incident information and your usage. If you need to hurry, you can try another available model, but some incidents occur across multiple models. You don't need to move on to billing, paid resets, or deleting authentication information just because of a capacity error.

Separate where you check: authentication for a 401, the chat's state for thread not found, and the usage screen for a limit notice. Recording the date, time, full text, model, and remaining balance when the error recurs will help the next investigation.

If a ChatGPT response stops partway, or requests are limited during heavy parallel use, see Fixing "Message delivery timed out" and telling it apart from "Too many requests".

If no error appears but Codex feels slow or heavy, you can narrow down the cause by symptom in Causes and fixes when Codex is slow or heavy.

Frequently asked questions

Will I still get this error if I have a Pro subscription?

Yes. In our environment, the same message appeared on a Pro subscription, and the official incident records do not exclude Pro. Having usage remaining in your plan and the model being able to accept processing at that moment are separate things.

Can buying credits or paying for a reset fix it?

There is no official explanation that this works. Additional credits and similar options are mechanisms for your own usage allowance, and they are a separate matter from server-side congestion. First check the usage screen to see whether you have actually reached a limit.

Why does it still happen after I switch to another model?

Several models can become congested or affected at the same time. There are official cases where the same message appeared across multiple models. This is not a failure of your PC or account, so check the incident information and wait a while.

Should I restart the app or start a new chat?

Because the cause is server-side congestion, restarting or starting a new chat usually won't resolve it. Try these only if the app itself is frozen, and before that, save your request, your changes, and any unfinished items.