[Cyber] Codex App: gpt-daybreak-blue-latest rate limited by unrelated org (40M TPM) — regression of #10057

Error: Codex error: Rate limit reached for gpt-daybreak-blue-latest (for limit gpt-5.6-sol) in organization
org-BOvpEHVcDPTe8h4lZnwMO5Ly on tokens per min (TPM): Limit 40000000, Used 40000000,

same

Confirming on a different platform/plan: macOS 27.0.0 (arm64), ChatGPT Pro, Codex CLI 0.146.0-alpha.9.2 + Desktop app 26.727.40816, Sign in with ChatGPT. Same org (org-BOvpEHVcDPTe8h4lZnwMO5Ly), same limit group (gpt-5.6-sol), same 40M TPM numbers.

Some controlled results that may help narrow this down.

gpt-5.6-sol succeeds from the same account, same auth, in the same minute that daybreak 429s.

Interleaved runs, identical trivial prompt:

| model | result |
|---|---|
| gpt-daybreak-blue / gpt-daybreak-blue-latest | intermittent 429, roughly half of attempts |
| gpt-5.6-sol | 8/8 success, never failed |

Since the error blames for limit gpt-5.6-sol, and gpt-5.6-sol itself is served fine against that same limit group seconds earlier and later, the reported bucket does not appear to be genuinely exhausted.

This is not an entitlement problem — the status codes separate cleanly.

| model | response |
|---|---|
| gpt-daybreak-red, gpt-5.5-cyber-preview | HTTP 400 invalid_request_error — "not supported when using Codex with a ChatGPT account" |
| gpt-daybreak-blue, gpt-daybreak-blue-latest | HTTP 429 rate_limit_exceeded, or succeeds |

Models I'm not approved for return a clean 400. Daybreak Blue (which I am approved for — chatgpt.com/cyber reports trusted access active) returns 429 or serves. So the 429 path is being taken for an entitled model.

The reported numbers look synthetic. Across every rejection, Used is exactly equal to Limit (40000000) to the digit, while Requested tracks real request size (17882, 18103, 20337, 20341, 32083, 35902, 59984 in the original report). retry-after comes back as 26–89ms, which is not a plausible TPM window rollover.

Not tier-related. gpt-5.6-sol completes on both service_tier: "priority" and "default". Daybreak advertises no service tiers so Codex omits the field and the server records "service_tier":"auto"; overriding the catalog via model_catalog_json to force priority on the daybreak entry still 429s.

Failed response object (from the WebSocket frame) resolves to the base model:

{"status":"failed","model":"gpt-5.6-sol","service_tier":"auto",
 "error":{"code":"rate_limit_exceeded", ...}}

Response ID: resp_0f7b2b794cb60b14016a7ad60aa2b8819183095396307adfa0 (2026-08-11 ~07:58 UTC)

Partial client-side mitigation, for anyone blocked right now: the default stream_max_retries = 5 gives up right at the boundary — I had a run succeed on retry 5. Raising it recovers a good fraction of these, at the cost of a long stall. Since the built-in openai provider id can't be overridden, it needs a clone:

model_provider = "openai-retry"

[model_providers.openai-retry]
name = "OpenAI (deep retry)"
base_url = "https://chatgpt.com/backend-api/codex"
wire_api = "responses"
requires_openai_auth = true
supports_websockets = true
request_max_retries = 10
stream_max_retries = 20
stream_idle_timeout_ms = 300_000

Note supports_websockets = true — omitting it silently downgrades the clone to SSE.

This doesn't fix it (still ~50% and up to ~3 minutes of retrying), and -m gpt-5.6-sol remains the reliable fallback.

Consistent with the resolution on #10057 — "a temporary service issue that caused the responses API to incorrectly report rate limit errors when it shouldn't."

same here

stream disconnected before completion: Rate limit reached for gpt-daybreak-blue-latest (for limit gpt-5.6-sol) in organization org-BOvpEHVcDPTe8h4lZnwMO5Ly on tokens per min (TPM): Limit 40000000, Used 40000000, Requested 43794. Please try again in 65ms. Visit https://platform.openai.com/account/rate-limits to learn more.

+1. Getting an error on org org-BOvpEHVcDPTe8h4lZnwMO5Ly as well.

Also seeing this:

stream disconnected before completion: Rate limit reached for gpt-daybreak-blue-latest (for limit gpt-5.6-sol) in organization org-BOvpEHVcDPTe8h4lZnwMO5Ly on tokens per min (TPM):

Limit 40000000, Used 40000000, Requested 24727. Please try again in 37ms. Visit https://platform.openai.com/account/rate-limits to learn more.

Independent reproduction on macOS 27.0 with the ChatGPT/Codex desktop app 26.803.61601 (build 6396) and Codex CLI 0.147.0, authenticated with ChatGPT (no API key).

• gpt-daybreak-blue-latest repeatedly fails, including on trivial prompts.
• Regular gpt-5.6-sol works from the same account/environment.
• The Daybreak failure is attributed to for limit gpt-5.6-sol with Limit 40000000, Used 40000000; observed requested values include 25,439, 25,444, and 37,409 tokens.
• The organization ID in the error is not the authenticated ChatGPT account/workspace ID.
• Lowering reasoning effort does not help.
• Daybreak advertises no Priority/Fast service tier, so Codex omits the configured service_tier = "priority"; that setting is not causing the failure.

This independently supports a service-side Daybreak routing/limiter regression rather than individual account consumption.

Feedback ID: 019ff0f6-3956-7992-aa56-9fe004ed02d6

+1:

stream disconnected before completion: Rate limit reached for gpt-daybreak-blue-latest (for limit gpt-5.6-sol) in organization org-BOvpEHVcDPTe8h4lZnwMO5Ly on tokens per min (TPM): Limit 40000000, Used 40000000, Requested 150048. Please try again in 225ms. Visit https://platform.openai.com/account/rate-limits to learn more.

Same error for me.

+1