Codex CLI 0.162.1 (macOS, Pro): ~50% of turns fail with "Selected model is at capacity" (server_overloaded) since Oct 8, on all models, with 99% weekly quota left #53531

What version of Codex CLI is running?

codex-cli 0.162.1 (npm)

What subscription do you have?

ChatGPT Pro

Which model were you using?

Mostly gpt-6.1-sol (reasoning: high). Also reproduced on gpt-6-sol (high), gpt-5.6-sol (high) and gpt-5.6-luna (low).

What platform is your computer?

macOS 26.3.1, arm64 (Apple Silicon)

What terminal emulator and version are you using (if applicable)?

iTerm2 3.6.9

Codex doctor report

20 ok, 0 fail. Auth: ChatGPT tokens (no API key). Responses WebSocket handshake OK (HTTP 101). Provider endpoints reachable. App-server daemon running (persistent mode).

What issue are you seeing?

Since 2026-10-08, roughly half of my Codex CLI turns end with:

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

The session files record "codex_error_info": "server_overloaded".

From my local session files and log database (Oct 8 to Oct 13):

• 97 failed turns. In the session files alone, 79 of 157 turns failed (50%). On some days the failure rate was above 70%.
• 56 failures happened mid-turn, after the model had already streamed output and run tools. The next sampling request in the same turn was then rejected.
• It happens on every model I tried (see above), so switching models via /model does not help.
• /status shows 99% weekly quota remaining. OpenAI Support confirmed my usage is ~1% and that this is not a rate limit.
• Context was not an issue: one failure happened with 94% of the context left.

Representative failures (UTC):

Feedback/session thread ID: 01a13067-9ab7-7621-94bf-495bde73dd84

What steps can reproduce the bug?

1. Log in to Codex CLI 0.162.1 with a ChatGPT Pro account.
2. Start codex with model gpt-6.1-sol (reasoning high), or any other model.
3. Give it an ordinary task, such as reading files, running a script or editing an image.
4. In about half of the turns, the turn aborts with "Selected model is at capacity" / server_overloaded, often after one or more tool calls have already succeeded.

Things that did not help:

• Switching models (4 different models).
• Routing the model WebSocket through a completely different US residential egress IP. I verified there was no HTTPS fallback, and it still failed.
• Running with --no-daemon to bypass the shared app-server daemon.
• Logging out and in, changing my password, enabling 2FA.
• Waiting 1h / 5h / 24h.

What is the expected behavior?

Turns should complete normally when my account has quota available. If capacity is genuinely constrained, Codex should retry with backoff instead of aborting the turn mid-way, and the problem should not persist for 6 days on a single account.

Additional information

This looks account-specific: dozens of colleagues on the same network never hit this error at the same moments. I have a full timestamped export of all 97 failures (thread and turn IDs only, no content) and can share it with maintainers. Possibly related: #50252, #53407, #28507.