Selected model is at capacity in Codex: not your usage limit, and 6 fixes in order

What does "Selected model is at capacity" mean in Codex?

It means OpenAI has run short of servers for the model you picked. Your account is fine. Codex prints the line from one error in its open source code, ServerOverloaded in codex-rs/protocol/src/error.rs, and the usage limit is a separate error with its own words. Read yours against this table, which we compiled from that file on 18 October 2026:

The tell is two words. A personal cap always says "usage limit", and a workspace cap says "credits" or "spend cap". The capacity line says none of them, and it never names a time, because nothing about your account changed. OpenAI's etraut-openai account, which answers Codex issues on GitHub, put it plainly on 12 May 2026: "If you receive this message, it does not mean that you are being rate limited or are out of usage. It means that we are running low on capacity to serve this model in your region."

OpenAI's API reports the same kind of shortage as HTTP 503 server_is_overloaded. Its error code guide gives the cause as "The requested model is temporarily overloaded." Codex users report the line in the desktop app, the CLI, the VS Code extension and Codex inside the ChatGPT desktop app. It reached ChatGPT once, too. On 9 July 2026 OpenAI posted an incident for users getting it "across multiple models", per Pulsetic's copy of the status feed.

How long does Codex stay at capacity?

A single capacity error usually clears in minutes, and the worst posted days ended within 3.4 hours. We read OpenAI's status history on 18 October 2026. Three incidents flagged capacity: the 16 June "Codex 'Selected Model is at Capacity' Error" at 196 minutes, the 9 July model selection incident at 58, and the 17 July "Codex 5.6-sol Experiencing Increased Server-Overload Errors" at 203. None has flagged it since.

The forum tells a longer story. Users posted waves of the error on 15 and 16 June, 28 to 30 July, 7 to 10 September and 3 to 12 October. The last three got no capacity notice. In the September wave, the only status note that mentioned Codex was a 25 minute incident about "unexpected usage limit resets". One July poster called GPT-5.6 Sol "almost continuously unavailable" for over 48 hours. A green status page doesn't clear Codex.

From 1 August to 16 October 2026, 13 incidents on OpenAI's status page had Codex in the title. They ran from 54 to 506 minutes, with a median of 95, and four went past three hours. We counted minutes from OpenAI's first posted update to its recovery notice. The history page starts on 20 July, so we found the June and July capacity incidents by search, and the 9 July one is read from Pulsetic's mirror.

How do you fix "Selected model is at capacity", in order?

Retry once, then switch models, then start a fresh chat. Codex has already retried before you see the line. Its config reference sets 4 retries for a failed request and 5 for a broken stream by default. Pressing Enter ten more times adds little.

1. Keep the prompt. Copy it before you reload anything. In a live session, users report that typing "continue" in the same thread often picks the task back up.
2. Retry once after a minute. Short spikes clear on their own, and one retry tells you if this is a blip.
3. Switch the model. In the CLI, run /model or start with codex -m gpt-6-luna. In the desktop app and ChatGPT Work, use the model control under the composer. GPT-5.5 left Codex with ChatGPT sign-in on 14 October 2026, so it's no longer a fallback.
4. Open a fresh chat with the same model. OpenAI staff suggested this on the forum in October. Run /new, or /compact first to carry a summary across.
5. Check status.openai.com, then /status. An incident naming Codex means wait. A clean page plus room left in your usage means the shortage is on your model and region.
6. Report it. Run /feedback from the failing chat and note the time with your time zone, which is what OpenAI staff asked for. If /feedback itself fails, staff pointed users to a support case instead.

Some fixes don't help. Reinstalling Codex, deleting ~/.codex or buying credits adds no servers. One September poster tried 4 IP addresses in 3 regions with no luck. If every model fails, OpenAI's own advice during the 25 September outage was "Login via API key will unblock access at this time." API key sessions bill at standard API rates, per OpenAI's Codex pricing page, so save that route for a real deadline.

Does a capacity error use up your Codex usage limit?

No. The capacity line is not a usage limit, and OpenAI's engineer said it doesn't mean you're "out of usage". You can check that in ten seconds. /status in a live CLI session shows the active model and your usage, and chatgpt.com/codex/settings/usage shows the same meter on the web. If the meter has room and the error persists, it's capacity. OpenAI hasn't said whether a turn that dies halfway is counted.

The usage limit runs on its own clock. OpenAI's Codex pricing page, read 18 October 2026, estimates 15 to 160 local messages per five hours on GPT-6.1 Sol for Plus and Standard Business. It gives 5 to 45 on GPT-6 Astra and 350 to 3,000 on GPT-6 Luna, and adds: "Pro plans currently have no five-hour limit." Weekly limits may apply on top. So on Pro, a wall that never says "usage limit" is capacity, not the weekly cap. Codex usage limit lists each plan's window and what its wall message offers.

Codex's third look-alike is "rate limit exceeded", which means requests came too fast. Unlike the capacity line, it is about your own request rate, so slowing down for a minute clears it on the same model.

Why does Codex hit capacity on a paid plan?

Because a plan buys an allowance, not reserved servers, and OpenAI rations capacity by model and region. The reports cluster around launches. Posters tied September's wave to the GPT-6 Astra launch. Many were on Pro 20x accounts, and one said a Pro 5x account kept working at the same time. In October, one poster traced the error to a background GPT-6 Luna request that names the chat. Staff replied that they "can't confirm the background Luna explanation from the thread alone."

The real fix is OpenAI's. On 15 October 2026 its staff wrote on the forum: "We've added capacity and shipped changes today to reduce these interruptions." A request for Codex to retry capacity errors with backoff, issue #22390, has been open since 12 May 2026. Until it ships, an unattended run stops at the first capacity error that outlasts the built in retries.

What can you use while Codex is at capacity?

Another vendor's model, for the part of the job that doesn't need your repo. A question about a stack trace, a plan for a refactor or a regex can go to Claude or Gemini while Codex waits. Paste a short summary of the task (the goal, what's decided, what failed) and carry on.

Whizi, which publishes this page, doesn't replace Codex. It has no developer API and doesn't touch your repository. It keeps GPT, Claude and Gemini in one chat: switching models mid conversation sends the thread to the new model (how switching works), and a generation that fails is refunded, not charged. Claude Sonnet and Claude Opus need Pro at $29.99 a month, and every plan is capped at 60 messages an hour. When the shortage is OpenAI's, pick a Claude or Gemini model.