Codex 'At Capacity' Error: October 2026 Outage Log and 4-Step Fix
If Codex shows "Selected model is at capacity", it is usually the server side running short on compute, not your network or your configuration. Check the official status page and /en/down/codex first to confirm there is no ongoing incident, then upgrade the CLI to the latest version, and only then look at your local network. This post collects the official incident records from late September 2026 through October 14, and gives you a checklist you can follow in order.
Which official Codex incidents were recorded in this period?
The OpenAI status page history shows more Codex-titled incidents after September 17, 2026 than the table below (there was one each on September 18 and 20). The table lists only the 8 incidents from September 25 onward, based on the status page titles. Times are as listed on the status page; check the page for the exact time zone.
| Date | Status page title (paraphrased) | Notes |
|---|---|---|
| Sep 25 | Issues with Codex | None |
| Oct 3 | Elevated error rates in ChatGPT and Codex | Some Codex remote control users had to re-pair their phones |
| Oct 6 | Elevated error rates in Codex Cloud | None |
| Oct 6 | Elevated error rates in ChatGPT, Codex, and the API (including the Agents API) | None |
| Oct 7 | Dot may be unable to create new threads | None |
| Oct 7 | Degraded performance in Codex and Work mode | The status page mentions slow responses and failed turns |
| Oct 13 | Degraded performance for one model in Codex and ChatGPT Work | None |
| Oct 14 | Degraded performance in Codex and ChatGPT Work | None |
The status page gives only a title and a short note. It does not state a root cause, and it does not always quote the exact error text. OpenAI has not said which of these incidents corresponds to "at capacity", so do not map them yourself.
How do I tell an official outage from my own network?
Work through these steps in order. It takes a few minutes.
- Open the OpenAI status page and look for any unresolved Codex or ChatGPT incident.
- Open cdnping's /en/down/codex and /en/status/openai. The first combines the official status page, API probes every 5 minutes, and user reports. The second probes the official API entry points from the Cloudflare edge. Note that cdnping only started collecting data on 2026-10-17, so it reflects the present and cannot serve as historical statistics for this period.
- If the official page and the probes both look normal but only you are seeing errors, check your local setup: egress IP, DNS, and whether your proxy is consistent. See Is ChatGPT down, or is it just your network? A three-step check and the Developer network self-check list (2026 edition).
- If the error says "at capacity", or shows a 429 or 5xx, and other people are reporting it too, it is most likely a server-side problem. Waiting and retrying works better than fiddling with your network.
For how to read the status page and what the incident levels mean, see How to read the OpenAI and Claude status pages. For what happened to other services in the same period, see AI service incidents roundup: OpenAI and Claude outages, September-October 2026.
Does upgrading the CLI help?
Somewhat. These are the overload-related changes in the official changelog:
- Codex CLI 0.161.0 (2026-10-07): the entry says Responses retries, and the fallback from WebSocket to HTTP, now follow the server's retry guidance, to reduce premature failures during overload. It also fixes incomplete history when resuming a thread.
- Codex CLI 0.162.0 (2026-10-08): for retryable Responses and WebSocket failures, the client follows the server's Retry-After guidance, and executor reconnects now include jitter.
In other words, older versions are more likely to fail outright when the server is overloaded. After upgrading, if you hit the same error, the client retries at the interval the server gives instead of giving up immediately. Upgrade the way you installed it (npm, a package manager, and so on); the exact command is in the official docs. After upgrading, run codex --version and confirm it is 0.162.0 or later.
Did OpenAI say it added compute?
I could not find any explicit statement about "added compute" or "automatic session recovery" on the official status page or in the changelog. A post on X says that a person in charge announced around October 15 that compute was being increased and automatic session recovery added, with the aim of reducing "Selected model is at capacity" errors. That claim has only this one source, and I could not verify it through official channels, so treat it as unconfirmed. The only changes I can confirm are the retry and recovery changes in the previous section.
What can I do while errors are happening?
- Do not hammer the retry button. A burst of requests just adds to the queue, and the newer CLI already backs off according to the server's guidance.
- If a long task is interrupted, use the CLI's resume-session feature to continue instead of starting over. The exact subcommand is in the official docs.
- Try again at a different time. Check the updates on each incident on the status page for recovery times.
- Remote control users may need to re-pair their phones after the October 3 incident.
- If your own connection is slow but the service is fine, use the cdnping home page to measure first-connection and round-trip times to each provider's API, then compare against AI API latency breakdown to see which stage is slow.
How do I keep tracking this?
- Official changes: see the Codex log in /en/changes, which records additions and removals line by line and has an RSS feed.
- Usage limits: running out of quota is a different message from a capacity error and needs separate handling. See How Codex usage limits work and Codex usage limit reset timeline.
- For setting a custom endpoint, see Custom API endpoints for Claude Code, Codex, and OpenCode.
Sources
- OpenAI status page incident history, checked on October 17, 2026
- ChatGPT changelog (Codex CLI 0.161.0, 0.162.0), checked on October 17, 2026
