Error messages

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

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

This is Codex telling you that OpenAI is out of room to run the model you picked, and you get it with usage left on the meter. That's what makes it so confusing. People on Pro plans with most of their weekly allowance untouched have been filing it on GitHub since at least April, and 228 issues in the openai/codex repo quote the exact line as of September 23, 2026.

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

We read it out of Codex's own source and the 0.156.1 release, and went through what OpenAI's engineers have said about it in those threads.

What OpenAI staff say it means

Eric Traut, who answers a lot of the Codex issues for OpenAI, has explained it twice. In April: "This isn't a rate limiting message", and "it's not specific to your account. This message indicates that we are out of capacity on this model." In May, on another thread, he was more specific: "It means that we are running low on capacity to serve this model in your region."

So your plan, your five-hour window and your weekly allowance aren't what tripped it. If you'd hit those, Codex has separate messages that start "You've hit your usage limit" and point you at an upgrade or at more credits.

Why Codex doesn't retry it for you

This part surprised us. In codex-rs/protocol/src/error.rs the line belongs to an error called ServerOverloaded, and the function that decides how long to wait before a retry returns nothing for it. That makes it terminal. Codex gives up on the turn at once and prints the line with a square bullet:

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

Compare the dropped-stream error and the server-error line, We’re currently experiencing high demand, which may cause temporary errors. Codex does retry both of those with backoff. The capacity one is handed straight back to you, which is why the message asks you to change model rather than wait. One user on the April thread asked why Codex doesn't queue the request instead, and the question is still sitting there.

What to do

Run /model, pick a different one, and send the prompt again. In codex exec pass -m with another model name.

Waiting also works, just more slowly. The people who reported it in April and May said it went away on its own, and Traut closed the main thread on June 16 with "This issue has been mitigated." A June report for GPT-5.5 was tied by a commenter to an incident on status.openai.com, so check the status page before you spend an afternoon debugging your setup.

When every model says it

September's reports look different. On September 7 several people filed issues saying GPT-6 Astra, GPT-5.6 and GPT-5.5 all failed at once, with only a mini model getting through. The most detailed one (issue #43337, 64 comments so far) comes from someone with two Pro 20x accounts: one fails every time with a minimal codex exec call, while the other works normally under comparable conditions. They saw it on Linux, macOS and Windows, and the minimal repro returns this:

{"type":"error","message":"Selected model is at capacity. Please try a different model."}

Nobody from OpenAI had replied on the two we read by September 23. If switching models does nothing and it follows your account from machine to machine, run /feedback right after a failure so the logs are uploaded, and link the upload on one of the open September issues. That reporter had already tried the CLI, the desktop app and three operating systems, so a reinstall is unlikely to be the answer.

The line that looks similar and isn't

Quota exceeded. Check your plan and billing details. comes from the same file, and it points at your plan and billing instead of OpenAI's servers. Codex doesn't retry that one either, and it's the case where changing model isn't the fix.

For the Claude Code version of a busy service, see what “Repeated 529 Overloaded errors” means.