Two post-resolution Codex failures are reported. Broad model decline remains unproven.
Current call: request and access problems are still reported on specific Codex paths; public evidence does not establish broad GPT-6.1 Sol capability decline. OpenAI marked its 13 October GPT-6.1 Sol Codex/Work performance issue resolved at 16:36 UTC. One Windows Pro account later reported capacity errors across seven models; a separate user diagnostic reported a server error tied to one saved reasoning item. The two reports describe different failures and do not establish a shared cause.
What the reports show
| Source and timing | Observed signal | What it does not establish |
|---|---|---|
| OpenAI Status, 13 Oct | OpenAI marked the GPT-6.1 Sol Codex/Work degraded-performance event resolved at 16:36 UTC. The listed surfaces include Codex Web, API, CLI, VS Code and ChatGPT Work. | The status record is aggregate. It does not show a particular account recovering or identify the root cause. |
| Codex issue #53559, 12–14 Oct | One Windows Pro user reports capacity messages and HTTP 503 errors across seven available models while 94% of the weekly allowance remained. Their 14 Oct snapshot reports 88 capacity errors and 55 HTTP 503s; the latest cited capacity error is 08:23 UTC. The author says Support confirmed one supplied request was server-overload and passed its usage check. | These are one account's logs and the author's summary of a support response, not an incident census. The issue page contains no later account result or proof of a broad quality change. |
| Codex issue #53807, 13–14 Oct | A separate user diagnostic reports one saved reasoning item failing before output on GPT-6.1 Sol: 4/4 default-tier and 2/2 priority tests. Requests with no saved history, with the other 145 items, or with the same item on GPT-6 Sol completed. | Small tests from one account isolate an input/history-specific request error; they do not establish population risk, backend cause or completed-answer quality. |
| Codex issue #53143, opened 11 Oct | An earlier user report describes a similar saved-reasoning error, with removal and GPT-6 Sol controls completing. Its tests began before OpenAI's 13 Oct status incident. | The author leaves backend attribution open. Similar symptoms do not prove the two cases share a cause or that the later status event caused them. |
Separate the failure modes
The Windows report is an account-level capacity signal across several models. The saved-history report is a request error tied to one particular input on GPT-6.1 Sol. Both stop the request before a completed answer can be judged. Neither report shows that the model returned a worse answer on a matched task.
The first saved-history report predates the status incident; the later replay was tested after the status page marked its event resolved. The timeline makes the distinction worth monitoring, but no provider log correlation explains whether the reports are related.
What to do next
- Check the current status page for the product surface and model you use.
- For a capacity or server error, note the time, plan/allowance display, client, model and surface; do not label the event a quality failure.
- For a saved-history failure, keep the original unchanged. If safe, try a small non-sensitive task in a fresh conversation using the same model and settings.
- For an answer that completes but seems worse, repeat the same task under the same surface, model, effort and tools; record correctness and correction burden separately from speed and allowance.
Sources
- OpenAI Status: degraded performance in GPT-6.1 Sol Codex and Work mode
- OpenAI Codex issue #53559: capacity and HTTP 503 report
- OpenAI Codex issue #53807: saved-reasoning replay report
- OpenAI Codex issue #53143: earlier saved-reasoning replay report
Return to the current verdict · Read how evidence is assessed