All Codex clients (Desktop app, codex exec, mcp-server) get 404 from chatgpt.com/backend-api/codex/responses since 2026-09-03 14:39 UTC (Windows, Pro)

[koniaes]

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

Codex App 26.901.1978.0 (Windows package OpenAI.Codex_26.901.1978.0_x64), bundled CLI codex-cli 0.153.0-alpha.5. Reproduced with the same binary through the Desktop app, codex exec and codex mcp-server (used from Claude Code via MCP).

What subscription do you have?

Pro (ChatGPT login)

What platform is your computer?

Windows 11 Home 10.0.26200, x64. Network egress from Spain (Cloudflare POP MAD).

What issue are you seeing?

Since 2026-09-03 14:39:25 UTC every turn fails with:

unexpected status 404 Not Found: Unknown error, url: https://chatgpt.com/backend-api/codex/responses, cf-ray: a355a9cdf90fec87-MAD

Same error from three independent clients that share the binary and the auth:

Client	Result
Codex Desktop (new and existing threads)	404 on every turn since 14:39:25 UTC; the rollouts before 14:27 UTC are clean
codex exec --sandbox read-only "…"	WebSocket wss://chatgpt.com/backend-api/codex/responses → 404 ×5 reconnects, then HTTPS fallback → 404 ×5
codex mcp-server (MCP tool codex)	404, six calls between 14:52 and 15:09 UTC

What I ruled out locally:

• codex login status → Logged in using ChatGPT; access token valid until 2026-09-08; account is pro.
• account/rateLimits/read over codex app-server works (Pro, weekly 24 % used), so account endpoints answer; only the responses route is 404.
• Unauthenticated POST https://chatgpt.com/backend-api/codex/responses → 404, while POST /backend-api/me → 403. The route itself is not answering.
• Model is gpt-5.6-sol (unchanged since Aug 29). Forcing gpt-5.2-codex → same 404.
• The binary (0.153.0-alpha.5, installed by the app at 00:05 UTC) was working: last successful MCP turn 13:03 UTC, last successful Desktop turn 14:27 UTC. The Desktop app was relaunched at 14:37 UTC; the first 404 is two minutes later.
• status.openai.com shows no incident as of 15:02 UTC.

Looks like the same outage as #42538, #42537, #42536, #42534, #42533 and #42532 (all opened today, macOS/Windows/VS Code, Plus and Pro).

What steps can reproduce the bug?

1. Open Codex Desktop, start a new thread, send any prompt → 404.
2. Or: codex exec --sandbox read-only --skip-git-repo-check "Reply exactly: PING_OK" → 404 after the reconnect attempts.
3. Or: start codex mcp-server and call the codex tool with any prompt → 404.

What is the expected behavior?

The model answers; backend-api/codex/responses is reachable (it was until 14:27 UTC with the same binary, token and config).

Additional information

• cf-ray ids seen (all -MAD): a35594b9db91547b, a35595f6bfe8ec96, a355970acb14d8ff, a355a076bfbfe8a4, a355a9cdf90fec87, a355ac86de34ec9a.
• codex exec also logs, before the 404 lines: ERROR rmcp::transport::worker: worker quit with fatal: Transport channel closed, when UnexpectedServerResponse("HTTP 404: ").
• Still failing at 15:09 UTC. A watcher retries every 10 minutes; I will add a comment with the recovery time when it comes back.

[github-actions]

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

• unexpected status 404 Not Found: Unknown error, url: https://chatgpt.com/backend-api/codex/responses, cf-ray: a35595eccb1ff781-MAD #42538
• ID 01a063cd-83f1-7df1-a261-6607f02cacbc #42532
• Codex Desktop 26.830.60940 — all models return 404 after upgrade on macOS #42537
• unexpected status 404 Not Found #42536
• Unexpected 404 error in codex #42534

Powered by Codex Action

[koniaes]

Recovered on my side: 2026-09-03 15:19 UTC. My probe at 15:09 UTC still got the 404; the next one at 15:19 UTC got a normal answer in ~10 s, and the MCP server answers again too. Same binary, token and config as during the failure, nothing changed locally. Outage window here: ~14:39 UTC → somewhere between 15:09 and 15:19 UTC (30–40 minutes). Correction to my earlier wording: this issue was created at 15:17:46 UTC, i.e. inside that recovery window, not after it.

[TeachDian]

https://status.openai.com/incidents/01M1KWEDH417T2CF44YYHZDFCR

i think we have the issues because of this

[koniaes]

Closing as a duplicate of #42534, which reported the same outage ~23 minutes earlier (14:54 UTC vs 15:17 UTC here) and already collects the recovery reports. The bot was right.

Leaving the timings here in case they help anyone correlating: outage started 14:39:25 UTC for me (Windows, Pro, MAD edge), recovered by 15:19 UTC — matching the status page incident "Elevated errors across ChatGPT and Codex" (investigating 14:58 UTC, mitigation applied 15:22 UTC). All three clients sharing the same binary and token failed identically (Desktop app, codex exec, codex mcp-server), and account/rateLimits/read kept answering throughout, so only the responses route was affected.