[Windows] Codex Desktop: repeated "Selected model is at capacity" (server_is_overloaded, NO_MORE_RETRY) with weekly usage at 1% - 171 failed turns over 3 days #53594

What happens

Since 2026-10-10, Codex Desktop on Windows repeatedly fails turns with:

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

The client-side error code is server_overloaded. The failures are intermittent but frequent: 171 failed turns over 3 days (2026-10-10: 82, 2026-10-12: 52, 2026-10-13: 37), spread over 14 threads. Long-running local workflows fail most often; retries (capacity_retry_automatic) and the automatic model switch (GPT-6.1 Sol -> GPT-6 Astra) both keep failing.

Environment

• Codex Desktop 26.1007.2314.0 (app build 26.1007.21434), Windows 11
• Bundled codex-cli 0.162.0-alpha.17.2 (the desktop app spawns this core for model traffic)
• Signed in with a ChatGPT account; model gpt-6.1-sol, reasoning effort xhigh, service_tier = default
• No image generation, no parallel sessions/devices

Evidence (captured from the model channel)

I captured the model transport with a local TLS-terminating proxy (GET /backend-api/codex/responses, WebSocket upgrade -> 101 Switching Protocols). The server returned this verbatim, 3 times in a row on 2026-10-13:

{"type":"error","error":{"type":"service_unavailable_error","code":"server_is_overloaded",
 "headers":{"x-retry-metadata":"NO_MORE_RETRY"},
 "message":"Our servers are currently overloaded. Please try again later.","param":null},
 "sequence_number":2}

each followed by a response.failed (e.g. resp_01868bcc51dd446d016acdd216987087d08423f761d74b5233), plus a separate server_error / HTTP 503 on the same channel.

Timestamps (UTC+8): 14:39:10, 14:39:21, 14:39:33 — with x-retry-metadata: NO_MORE_RETRY.

Quota is not the cause

On the same WebSocket connection, the server pushes rate-limit frames:

{"type":"codex.rate_limits","plan_type":"prolite",
 "rate_limits":{"allowed":true,"limit_reached":false,
 "primary":{"used_percent":1,"window_minutes":10080},"secondary":null},
 "credits":{"has_credits":false,"unlimited":false,"balance":"0"}}

allowed=true, limit_reached=false, weekly window at 1% (peak 14% across all samples), no credits, no spend control triggered. A successful turn on the same channel minutes later returns response.completed, so connectivity and auth are fine.

The account only exposes a weekly window (secondary: null) - there is no 5-hour window to look at.

What I'd like checked

This looks like capacity/routing decided server-side per account, not client-side throttling. Could someone check server-side logs for these thread IDs and confirm whether gpt-6.1-sol is being rejected for this account specifically?

• 01a132fb-19f3-7d73-b965-f981d67a4eb3 - 23 failures, 2026-10-13 09:59-14:33 (UTC+8)
• 01a11e8a-d742-7d12-a4df-4e7b18466573 - 30 failures, 2026-10-10 16:29 - 2026-10-12 20:03 (UTC+8)
• 01a12dd9-8489-7ca0-9aa5-7532ec9068ac - 19 failures, 2026-10-12 10:05 - 2026-10-13 09:33 (UTC+8)

Account locators (also raised through OpenAI support): creator_account_id 4e03fc68-6abc-4e6b-a154-baed7caa6573, user id user-aNs9nnp10DQm401Tlk5alb9x.

Related issues

Same symptom class, several explicitly account-specific:

• Account-specific capacity errors across Codex models despite a fully available weekly allowance #43337 Account-specific capacity errors across Codex models despite a fully available weekly allowance
• Account-specific server_is_overloaded on GPT-5.6 Sol: one Plus account fails while two other Plus accounts work on same machine/network #46133 Account-specific server_is_overloaded on GPT-5.6 Sol: one Plus account fails while two other Plus accounts work on same machine/network
• [Bug] "Selected model is at capacity" fires in bursts across all models on an APAC/Japan Pro account #47168 "Selected model is at capacity" fires in bursts across all models on an APAC/Japan Pro account
• Account-specific server_overloaded / capacity on Pro despite usage remaining (thread 01a1337b-ddfb-7013-ba44-2716c02d2789) #53555 Account-specific server_overloaded / capacity on Pro despite usage remaining
• [Windows 10 19045] GPT-6.1 Sol: persistent "at capacity" / HTTP 503 with 94% quota remaining (Oct 12-13), corroborating #52812 / #53407 #53559 [Windows] GPT-6.1 Sol: persistent "at capacity" / HTTP 503 with high quota remaining (Oct 12-13)
• Codex long-running tasks repeatedly hit server_overloaded / At Capacity despite ~70% usage remaining #52394 Long-running tasks repeatedly hit server_overloaded / At Capacity despite ~70% usage remaining
• Codex "Selected model is at capacity" / server_overloaded on trivial first turn (Web + Windows app) #52812 "Selected model is at capacity" / server_overloaded on trivial first turn (Web + Windows app)
• [Windows][Codex Desktop] Long-running Local workflows intermittently fail with server_overloaded #52833 [Windows][Codex Desktop] Long-running Local workflows intermittently fail with server_overloaded
• Codex Desktop: All models report "Selected model is at capacity" on one Pro account, while another account works on the same Mac #52894 All models report "Selected model is at capacity" on one Pro account, while another account works
• Codex App repeatedly shows "Selected model is at capacity" despite healthy connectivity #45835 Codex App repeatedly shows "Selected model is at capacity" despite healthy connectivity

Note on evidence

The frames above are copied verbatim from a HAR captured on the model WebSocket channel; the full capture (400 KB, or 28 KB zipped) is available and can be attached if useful.