Selected Model Is at Capacity: What to Do

What does "Selected model is at capacity" mean?

"Selected model is at capacity. Please try a different model." means OpenAI is short of capacity to serve that model right now. It is a server-side failure in Codex, the OpenAI coding agent. It does not mean your plan allowance has run out.

Try a different model or wait, and check the status page. If your message says You’ve hit your usage limit instead, you have a different problem, so go to what to do when you reach a ChatGPT limit.

The clearest statement comes from a Codex maintainer at OpenAI, in a GitHub comment dated 12 May 2026:

If you receive this message, it does not mean that you are being rate limited or are out of usage.

The same comment continues: "It means that we are running low on capacity to serve this model in your region." This is a comment on OpenAI's own repository. We found no OpenAI help article that covers this message, so treat the comment as the best statement available. The maintainer closed that issue on 16 June 2026 with the words "This issue has been mitigated". The message is still in the code, so it can return.

Where does the message come from?

The wording is fixed text in the Codex source code. It is the ServerOverloaded error in codex-rs/protocol/src/error.rs. We read this file on the main branch on 6 November 2026, and upstream code moves.

#[error("Selected model is at capacity. Please try a different model.")]
ServerOverloaded,

This error variant has no fields, so its text never changes. It is the same on every plan, and it never names the model. It also cannot show a region or an expected wait, so the message will not tell you how long the problem will last.

Codex shows this text when a response has HTTP (hypertext transfer protocol) status 503 and the error code server_is_overloaded. That mapping is in codex-rs/codex-api/src/api_bridge.rs. The same file has a second path that turns an internal "server overloaded" error into this message, so a 503 with that code is not the only route to it.

Is it a capacity error or a usage limit?

Read the exact words on your screen. The two errors come from different responses, and they need different fixes.

• Capacity error. The text is Selected model is at capacity. Please try a different model. It comes from an HTTP 503 response. The shortage is on OpenAI's side.
• Usage limit error. Most versions of the text start with You’ve hit your usage limit. Workspace versions can start with Your workspace is out of credits or You hit your spend cap instead. Codex builds this error only from an HTTP 429 response whose body has the type usage_limit_reached. Your allowance, your workspace credits or your spend cap is the thing that ran out.

The rest of the usage limit text depends on your plan. Plus users are told to upgrade to Pro. Free and Go users are told to upgrade to Plus. Pro users are sent to a page to buy more credits. Enterprise and Edu users get only the first sentence and a line about when to try again.

Business and Team users see a message that starts like this:

You’ve hit your usage limit. To get more access now, send a request to your admin

That line belongs to the usage limit error. The at-capacity text has no admin line, so if you see the admin line, the steps in this guide are the wrong ones. This wording is in the Codex source. We did not find it on the pricing page or in the Codex help article.

One usage limit message can confuse you, because it also talks about models: You’ve hit your usage limit for a named limit, then Switch to another model now. It is still a usage limit, because the first words say so.

Codex also has other service messages in the same source file. They are separate errors, so do not assume the advice here covers them:

• Policy-restricted capacity is unavailable. Please try again later.
• Flex capacity unavailable.
• We’re currently experiencing high demand, which may cause temporary errors.

What should you do when a model is at capacity?

Work through these steps in order.

1. Confirm the exact text. Only Selected model is at capacity is this error.
2. Check the OpenAI status page. It lists Codex components: Codex Web, Codex API, CLI (command-line interface) and the VS Code extension.
3. Run /status in the Codex CLI to see which model is active, because the error does not name it.
4. Try a different model with /model, as the next section shows.
5. If a different model fails too, stop switching. Wait, then try again.
6. If it continues, report it with /feedback.

A healthy status page does not prove the problem is on your side. The page says its availability figures are aggregated across tiers, models and error types. So a shortage for one model in one region may not appear there as an incident. Use the page for live state.

How do you switch model in Codex?

Inside a Codex CLI session, type this and press Enter:

/model

A list of models opens. It shows only models your account and workspace can use. Pick one, and Codex confirms the change in the transcript. Then check it:

/status

/status should now show the new model as the active model. It also shows your approval policy and current token usage.

If your session closed, go back to the project folder and resume the most recent chat from that folder, then use /model again:

codex resume --last

To start a new session on a chosen model, use the --model flag. The name below is the example from the OpenAI docs. Replace it with a model from your /model list, and do not pick the model that just failed. If your list offers more than one GPT-6 tier and you are not sure which to fall back to, our comparison of GPT-6 Astra, Sol and Luna explains what each one is for.

codex --model gpt-6.1-sol

codex exec accepts the same flag for one non-interactive run. That is the place to change it if a script fails with this error.

The default model lives in ~/.codex/config.toml, and a trusted project can set its own in .codex/config.toml:

model = "gpt-6.1-sol"

CLI flags and --config overrides take priority over these files. So if /status shows a model you did not expect, check these layers. All of this comes from the Codex command reference and the Codex config docs.

These commands are for the CLI. The OpenAI docs say ChatGPT web has its own command menu and does not expose the CLI command set, so do not expect /model to work there.

Does switching model actually clear the error?

It might, but nothing promises it. The error text says to try a different model. The maintainer's comment describes a shortage for "this model in your region", which fits with another model still working. No OpenAI page we opened says that switching reliably clears the error, and we have no success rate to give you.

So treat a switch as a test. If the new model answers, continue your work. If the new model gives the same error, the shortage may be wider than one model, and more switching will not teach you anything.

A switch has one cost to think about. OpenAI says Codex usage depends on the model, along with where the task runs, task complexity, context, reasoning, speed and tools. A different model can therefore use your allowance at a different rate. For how the allowance works, read how Codex works with a ChatGPT subscription.

Does a failed attempt use any of your allowance?

OpenAI does not say. The Codex help article says "Model usage counts toward your normal Codex usage limits and applicable credits or billing." It says nothing about failed or errored requests. The API error codes page does not answer the question either.

The maintainer's comment says the message does not mean you are out of usage. That is a statement about the cause of the error. It does not say whether the attempt was counted.

You can check this for your own account. Open your usage page at chatgpt.com/settings/usage before and after a failed attempt, and compare the two. /status also shows token usage for the current session. Do not assume a failed attempt is free, and do not assume it is counted.

How long should you wait before you retry?

OpenAI publishes no fixed interval for this error, and we will not guess one. The API error codes page describes 503 as "Model temporarily overloaded". Its advice is to follow the Retry-After header when it is present, and to increase the delay between retries when it is missing. That page is about the API, but it is the closest written rule from OpenAI.

The Codex source also follows the server's delay for this error. Timeouts and other temporary errors fall back to a local delay when the server sends none, but ServerOverloaded does not. So if the server sends no delay, Codex has nothing to wait on, and you should expect to retry by hand.

If a script calls Codex, make each wait longer than the last one. A fast loop only sends more requests to a model that is already short of capacity.

Does the same wording appear in ChatGPT itself?

We found no primary evidence that it does. The string we verified is in the Codex client source. OpenAI's ChatGPT error troubleshooting article does not contain the words "at capacity", "Selected model" or "try a different model". Its closest entry is the generic Something went wrong. message, which it calls a general error that may point to a temporary server issue or a problem with your local setup.

The status page does list "Codex in ChatGPT Desktop" as a ChatGPT component. That may explain why some people connect this message with ChatGPT, but no OpenAI page confirms it. If you saw the message in the ChatGPT app, note where you were working when you report it.

For a generic error in ChatGPT, that article suggests checking the status page, refreshing or starting a new chat, and clearing your cache and cookies, among other steps.

What if your plan limit really did run out?

Then capacity steps will not help. OpenAI's help article says you may be able to add credits, use an available reset, upgrade, or wait until the displayed reset time. It also says "OpenAI Support does not reset ChatGPT or Codex usage limits." and "Starting a new chat does not reset your usage limits." On an Enterprise workspace billed by tokens, you follow your organization's process for requesting more usage.

For the reset times, read ChatGPT usage limits and when they reset. If you are thinking about a paid plan because of real limits, compare the ChatGPT Plus price and limits first.

How do you report the error to OpenAI?

Type /feedback in the Codex CLI and press Enter, then follow the prompts. It sends logs to the Codex maintainers.

You can also open an issue in the openai/codex repository. The old issue is locked, and the maintainer asked people who see the problem again to open a new one. Write down the exact message, the active model from /status, the time, your region, and whether you used the CLI, the app or the VS Code extension. Those details matter because the cause is tied to a model and a region.

FAQ

Does "Selected model is at capacity" mean I ran out of usage?

No. A Codex maintainer at OpenAI wrote that the message does not mean you are rate limited or out of usage. It means OpenAI is running low on capacity to serve that model in your region. A real usage limit has different text. Most versions start with You’ve hit your usage limit, and workspace versions can start with Your workspace is out of credits or You hit your spend cap instead.

Will switching to a different model fix the capacity error?

It may. The error text suggests it, and the stated cause is capacity for one model in one region. No OpenAI page promises that it works. In the Codex CLI, type /model, pick another model, and confirm the change with /status. If the second model fails the same way, wait and try again later.

Does a failed Codex attempt count toward my limit?

OpenAI's help article and API error codes page do not say. The help article says only that model usage counts toward your normal limits and applicable credits or billing. To check your own account, compare chatgpt.com/settings/usage before and after a failed attempt.

My error says to send a request to my admin. Is that the capacity error?

No. To get more access now, send a request to your admin is part of the usage limit error for Business and Team plans. Codex builds that error from an HTTP 429 response with the type usage_limit_reached. The at-capacity error comes from a different response and has no admin line.

Will a more expensive plan stop this error?

We found nothing from OpenAI that says so. The Codex pricing page has no wording about this error, and the maintainer's explanation is about model capacity in a region. Do not upgrade only because of this message. If you want to compare plans for other reasons, read which ChatGPT plan you need.