Newly created ChatGPT Free account fails Codex account warmup with HTTP 404

[Jiangxb8]

Client or integration

Codex CLI

Provider or upstream service

OpenAI ChatGPT / Codex

OpenCodex version

2.47.0

Endpoint or capability

ChatGPT OAuth account warmup / Codex responses endpoint

Current behaviour

I already have two Codex accounts configured in OpenCodex.

When I try to add a third ChatGPT account through OpenCodex GUI using OAuth login, the browser authentication appears to complete successfully, but OpenCodex fails during the account warmup step.

The error shown in the GUI is:

Codex account warmup failed. Reauthenticate the account and try again. (http_status:404)

The third account was newly created today. It is currently a ChatGPT Free account with no paid subscription.

The account itself works normally on the ChatGPT website: I can sign in successfully and start normal conversations.

Re-authenticating the account produces the same HTTP 404 warmup error.

Expected behaviour

After a successful ChatGPT OAuth login, OpenCodex should complete the Codex account warmup successfully and add the account to the account pool.

The account is a valid ChatGPT Free account and can sign in and use ChatGPT normally in the browser.

According to OpenAI's current documentation, Codex is available across ChatGPT plans, including the Free plan, although usage limits differ by plan.

Minimal redacted request or reproduction

Steps to reproduce:

1. Open OpenCodex GUI.
2. Go to the Codex account pool/account management page.
3. Select "Add Codex Account".
4. Enter a new account ID.
5. Select "OAuth Login".
6. Sign in with a newly created ChatGPT Free account.
7. Complete the ChatGPT OAuth authentication successfully.
8. Return to OpenCodex.
9. OpenCodex starts the Codex account warmup.
10. The warmup fails with HTTP 404.

Observed error:

`Codex account warmup failed. Reauthenticate the account and try again. (http_status:404)`

Additional details:

- OpenCodex version: 2.47.0
- This is the third Codex account I am trying to add.
- The first two accounts are already configured.
- The failing account was created today.
- The failing account has no paid subscription (ChatGPT Free).
- The account can log in and chat normally at chatgpt.com.
- Re-authentication does not resolve the problem.

Actual response or error

Codex account warmup failed. Reauthenticate the account and try again. (http_status:404)

Upstream documentation

https://help.openai.com/en/articles/11369540-using-codex-with-your-chatgpt-plan

Suggested mapping or implementation notes

It may be useful to check whether the Codex account warmup endpoint or model selection behaves differently for newly created ChatGPT Free accounts.

Since OAuth authentication succeeds but the subsequent warmup request returns HTTP 404, this may be related to the warmup request itself rather than authentication.

It may also be worth logging the redacted warmup request URL/model and response body when warmup fails, so that 401/403 authentication failures can be distinguished from 404 endpoint/model/account-provisioning failures.

Additional context and attachments

The affected account was created today and currently has no paid ChatGPT subscription.

However, the account works normally in the ChatGPT web interface and can create normal chat conversations.

The problem only occurs when OpenCodex performs the account warmup after OAuth login.

I have attached a screenshot of the OpenCodex error dialog.

If useful, I can provide additional OpenCodex debug logs with account/token information redacted.

Checks

• I searched existing provider and compatibility issues.
• The request and response were redacted.
• The expected behaviour is based on an upstream specification or a concrete client requirement.

[Ingwannu]

Thanks for the detailed report. Current OpenAI pricing documentation includes Codex on Free, so I would not diagnose this as "Free accounts are unsupported": https://developers.openai.com/codex/pricing/ (checked 2026-09-09).

On current dev 57077ca326, src/codex/warmup.ts sends a separate inference request after OAuth, defaults to gpt-5.4-mini, and only tries its fallback on HTTP 400, not 404. verifyCodexAccountWarmup then presents every warmup failure as a reauthentication error. This confirms a misleading diagnostic path; it does not establish whether your 404 is model access, account provisioning, or an endpoint issue. The account's ability to chat on the website does not by itself prove that this exact Codex inference request succeeds.

Please report whether the same account can complete one ordinary turn in the official Codex client, and the client version/model plus a redacted status/error code if it fails. If you can retest OCX 2.49.0, report that result separately; an upgrade is not claimed to fix this. Do not post tokens, auth.json, account identifiers, cookies, or full debug dumps. I will not bypass successful inference validation or mark a 404 as a valid account merely to make registration succeed.

[Jiangxb8]

Thanks for the clarification.

I tested the same account in the official Codex client first. It was able to complete a normal turn successfully using the gpt-5.6-luna model.

I then retested adding the same account in OpenCodex 2.49.0. OAuth login still completes successfully, but the warmup fails with the same error:

Codex account warmup failed. Reauthenticate the account and try again. (http_status:404)

Current results:

• Official Codex client: successful
• Working model: gpt-5.6-luna
• OpenCodex version: 2.49.0
• OpenCodex warmup: still fails with HTTP 404

One additional observation may be relevant: in the official Codex app, if I select gpt-5.5 with this account, the client reports that this model is not supported.

Since OpenCodex warmup currently defaults to gpt-5.4-mini, I wonder whether this newly created Free account may also not have access to that model. If so, the 404 may be related to model availability/entitlement rather than OAuth authentication itself.

I cannot confirm this yet, but the model-access difference seems potentially relevant.

[lidge-jun]

리뷰 · 우선순위 64 / 80

증상은 OAuth는 성공하는데 Codex 계정 warmup만 HTTP 404로 실패하고, GUI가 “다시 로그인하라”고 말하는 것이다. 보고자는 ChatGPT Free·당일 생성 계정이고, 웹 ChatGPT 대화는 된다. Ingwannu 코멘트 이후 후속으로 공식 Codex 클라에서는 gpt-5.6-luna로 정상 턴이 되고, OpenCodex 2.49.0에서도 같은 404가 남으며, 공식 앱에서 이 계정으로 gpt-5.5는 “지원되지 않음”이라고 한다. Free 미지원으로 단정할 근거는 없고(OpenAI Codex pricing에 Free 포함), 모델 entitlement / warmup 진단 쪽으로 무게가 옮겨졌다.

현재 dev HEAD 3b4d8c439 코드 경로를 보면 일치한다. src/codex/warmup.ts는 https://chatgpt.com/backend-api/codex/responses로 별도 추론을 보내고, 기본 모델은 gpt-5.4-mini, 폴백 목록은 ["gpt-5.5"]뿐이다. warmCodexAccount는 HTTP 400일 때만 폴백을 시도하고, 404는 즉시 throw한다. src/codex/auth-api.ts의 verifyCodexAccountWarmup은 실패 원인을 가리지 않고 401 + “Reauthenticate the account and try again” + reason(예: http_status:404)으로 포장한다. 그래서 OAuth가 살아 있어도 UI는 재인증을 강요하고, 404가 모델/프로비저닝 문제인지 구분하기 어렵다.

관련으로 “Free 계정 warmup 404를 고친다”는 오픈 PR은 보이지 않는다. #3848 등은 할당량 소진 등록 유예 쪽이고 이 404와는 다르다. 재현에 필요한 다음 정보는 (1) warmup이 실제로 친 모델 id, (2) 404 응답의 redacted URL/코드만(토큰·auth.json 금지), (3) gpt-5.4-mini를 공식 Codex에서 직접 골랐을 때의 결과다. 보고자 가설(“Free가 mini/5.5에 없고 luna만 된다”)이 맞다면, 폴백을 400에만 묶은 현재 로직이 바로 막는다.

경로/심볼 - src/codex/warmup.ts DEFAULT_MODEL / FALLBACK_MODELS / warmCodexAccount — 404는 폴백하지 않음. 보고자 계정에서 공식 클라가 거절하는 gpt-5.5가 유일한 폴백인 점도 위험하다.

경로/심볼 - src/codex/auth-api.ts verifyCodexAccountWarmup — 모든 warmup 실패를 재인증 메시지로 통일. http_status:404는 재로그인으로 안 고쳐질 가능성이 크다.

경로/심볼 - 오픈 PR — 이 이슈를 직접 고치는 PR 없음. 진단 개선 + Free-safe 모델 폴백이 별도 패치 후보.

메인테이너의 판단이 필요한 지점

• 404에서도 모델 폴백을 시도할지(과도한 청구/노이즈 vs 등록 성공률)
• warmup 기본/폴백 목록에 gpt-5.6-luna 등 Free에서 검증된 모델을 넣을지
• UI 문구를 http_status일 때 “재인증”이 아니라 “모델/프로비저닝 확인”으로 나눌지
• 성공한 추론 검증을 완화하지 않는다는 Ingwannu 원칙을 유지한 채 진단만 먼저 고칠지

너의 추천
이슈는 열어 둔다. 다음 패치 후보: (1) warmup 실패 메시지를 status별로 분리하고 redacted 모델·status를 GUI/reason에 노출, (2) 404에도 안전한 폴백 체인 시도(최소 gpt-5.6-luna 등 보고자가 공식 클라에서 성공한 모델), (3) gpt-5.5만 폴백인 현재 목록 재검토. 보고자에게는 토큰 없이 “공식 Codex에서 gpt-5.4-mini 선택 결과”만 추가 부탁하면 충분하다. 등록 검증 자체를 끄거나 404를 성공으로 취급하지는 않는다.

이 댓글은 grok-bot이 작성했습니다

[lidge-jun]

Fixed on dev via #4230 at ad36a59. Warmup now retries Free-provisioned models (including gpt-5.6-luna) on HTTP 404/400 instead of blaming a fresh credential.