Persistent Codex “Selected model is at capacity” / HTTP 503 after usage reset on macOS #53589

What version of the Codex App are you using (From “About Codex” dialog)?

26.1007.30912

What subscription do you have?

Paid ChatGPT/Codex account

What platform is your computer?

Two Macs

What issue are you seeing?

Codex repeatedly returns:
Selected model is at capacity. Please try a different model.
The issue continues even after the weekly Work/Codex usage window reset. The Usage page shows approximately 4% remaining and the weekly window is not marked as reached. Logging out of all accounts and signing in again did not resolve the problem. Switching models has also not reliably resolved it.
The issue is intermittent. Some requests may eventually succeed after changing the network route or retrying, but other requests fail again. This makes long-running Codex tasks unreliable.
Local evidence
I performed a read-only inspection of the active Codex SQLite log on macOS. It contains:

• 14 unambiguous 503 Service Unavailable entries from codex_core::responses_retry
• Capacity-related records in session turn and session handler components
• Retry-control events
• Two bursts on October 13, 2026: 01:53–02:09 UTC and 03:24–03:26 UTC
• No unambiguous HTTP 429 entries
• No clear evidence that a local semaphore, queue, or concurrency limit caused the failures
  A fresh model-selection test also showed a mismatch: GPT‑6.1 Sol was requested, but the task metadata reported gpt-5.5 with high reasoning. The actual backend-served model could not be independently verified.
  Questions
  Could the team please investigate:

1. Whether these requests are being routed to an overloaded or unstable backend pool
2. Whether the account has an account-specific admission or routing issue despite remaining usage
3. Whether the client is silently falling back to a different model
4. Whether server_overloaded / HTTP 503 can be exposed clearly instead of being reported only as model capacity
5. Whether long-running tasks can retry or resume safely without losing context or repeating completed tool calls
   Public reports with similar symptoms include Selected model is at capacity (server_overloaded) on ChatGPT Pro at 4.0% weekly usage — 41 failures, not a usage limit, Fast/priority tier also affected #46798, Account-specific server_overloaded / capacity errors: Plus account 0/11 while Free account 11/11 on same machine, network, CLI, and model #46103, Selected model is at capacity. Please try a different model. No matter how I swtich model, reset, reinstall, logout/login #45653, and Codex "Selected model is at capacity" / server_overloaded on trivial first turn (Web + Windows app) #52812. Please correlate the timestamps above with backend request history if possible.
   This report intentionally omits private account identifiers, IP addresses, tokens, and session contents.

What steps can reproduce the bug?

What is the expected behavior?

No response

Additional information

No response