"Selected Model Is at Capacity" ChatGPT Error: How to Fix It

"Selected model is at capacity. Please try a different model." is the error ChatGPT and Codex throw when the model you picked has no free serving capacity in your region at that moment. It is not a rate limit, it is not your account, and it is not something you can fix by reinstalling or paying more. This guide covers what the error means, when it started, why it got dramatically worse after the GPT-6 Astra launch, what users around the world are saying about it, and every fix that people actually report working (including the VPS region trick, with honest caveats).

Quick answer: “Selected model is at capacity” means the specific model cannot accept your request right now. It usually clears when you wait and retry or choose another available model. It is different from a usage-limit message.

Fix it in this order

1. Check OpenAI Status for an active ChatGPT or Codex incident.
2. Wait 5–15 minutes, then retry once. Avoid repeatedly resending the same request.
3. Switch models in the model picker. Try a sibling or lighter reasoning model; capacity is model-specific.
4. If Codex stopped mid-task, send continue in the same thread after capacity returns so you keep the existing context.
5. Start a new chat or session if only one conversation is failing.
6. If every model fails, refresh the app, try another browser or network, and contact OpenAI Support with the model, timestamp, app version, and region.

These steps apply to ChatGPT and Codex. An API response such as 429 or an explicit usage-limit message is a separate problem and needs API quota or rate-limit troubleshooting.

What the "Selected model is at capacity" error actually means

OpenAI has never published a help center article for this exact message, but an OpenAI contributor did explain it directly on GitHub. Responding to issue #17014 on April 7, 2026, etraut-openai wrote:

"This isn't a rate limiting message. That is, it's not specific to your account. This message indicates that we are out of capacity on this model."

That one sentence rules out most of what people try first. Your usage bar is not the problem. Your Pro or Plus plan is not the problem. Codex is not broken. OpenAI simply ran out of serving slots for that model at that moment, and the frontend told you to pick a different one.

The error appears in three places:

• ChatGPT (web and desktop app) when you send a message on an affected model
• Codex app and Codex CLI mid-task, which is the worst version because it interrupts running work
• Codex IDE extensions (VS Code, JetBrains) in the same way as the app

The ChatGPT version tends to hit before your message sends. The Codex version tends to hit while an agent is mid-edit, which is why most of the anger online comes from Codex users.

When the error started, and why it got worse

The message has been around since at least spring 2026, but it went from occasional to constant in the last two weeks.

April 2026: first reports on GPT-5.4 and GPT-5.3-Codex

The first GitHub issue landed April 7, 2026 against GPT-5.4. On April 23 a wave of comments hit the same thread within minutes of each other, from users on Codex desktop, Codex CLI, and third-party editors, all seeing the error on GPT-5.3-Codex and GPT-5.4 during European working hours. One user summed up the frustration: "You should stop new requests rather than cutting off people's work halfway."

June 16, 2026: the first official incident

On June 16, OpenAI acknowledged it publicly. Vaibhav Srivastav from OpenAI posted on X: "Some Codex users are seeing a 'Selected Model is at Capacity' error. We're working to resolve it as quickly as possible." The status page incident listed ChatGPT and four Codex components as affected, was identified at 10:32 AM, and was marked resolved at 1:48 PM. About three hours of degraded service. OpenAI staff member VeitB then opened a forum thread asking anyone still affected to report in. Users on GPT-5.5 High, xHigh, and Medium kept reporting in for weeks.

July 2026: GPT-5.6 Sol becomes the main offender

By mid-July the complaints shifted to GPT-5.6 Sol. On July 29 a Pro 20X user opened a thread titled "GPT-5.6 Sol repeatedly hits 'Selected model is at capacity' in Codex Desktop," reporting it across Codex Desktop, JetBrains, and OpenCode on Sol, Sol Max, and Sol Ultra. OpenAI Support asked for their device, plan, and region or country, which is a hint about how capacity is allocated.

August 31 to September 2, 2026: the spike

Starting August 31, at least six new GitHub issues were filed in three days (#41790, #41798, #41805, #41808, #41810, #42169), nearly all naming GPT-5.6 Sol. The reporter of #41790, a ChatGPT Pro subscriber, hit the error three times inside one 15-minute task with plenty of usage remaining. Codex now shows an automatic retry countdown when this happens (117 seconds in that report), and users say the wait grows with each occurrence and the error often returns within minutes of the retry succeeding.

September 3, 2026: GPT-6 Astra launches, and the error hits GPT-6 too

OpenAI released GPT-6 Astra on September 3, first to a limited set of organizations, then to all Pro, Enterprise, and Business Premium users in ChatGPT and Codex, with Plus users waiting days for access. Everyone who had been saving usage for launch day tried it at once.

Within two days the capacity error was showing up on Astra itself. By September 7 and 8, Chinese developer forum LINUX DO had multiple threads titled things like "gpt-6-astra keeps saying 'Selected model is at capacity'" and "codex gpt-6-astra xhigh error: blown up." One reply in the second thread, translated: "Wait for the official fix, lots of people on X have this problem." Another: "Tibo is playing dead right now, it's infuriating." (Tibo is the Codex product lead.)

So as of this week the error affects the entire top of the lineup: GPT-5.6 Sol, GPT-5.6 Terra, GPT-5.6 Luna, and GPT-6 Astra, with the high and xhigh reasoning tiers hit hardest.

What users are saying on X and forums (translated)

The complaints are global, and the non-English ones are often more useful because they include what people tried.

Japan. @writer_china, an AI content creator, posted on June 9 (translated from Japanese):

"Error: 'Selected model is at capacity. Please try a different model.' If you don't check in now and then, the task just stalls. I showed ChatGPT the error screen and what I was doing. Cause: not a billing shortfall, not a permissions problem. GPT-5.5 (Codex) processing slots are temporarily full. Image generation with Image2 is especially GPU-heavy and prone to this during congestion. Fixes: 1) wait 5 to 15 minutes and rerun (most recommended), 2) run the same thing in a new chat, 3) switch to a model other than GPT-5.5, 4) wait longer for OpenAI's congestion to clear. Conclusion: this is not a 'pay money and it goes away' problem. It's OpenAI being congested and unable to process right now."

Malaysia. Colin Charles on July 12: "gpt-5.6-sol medium: ⚠ Selected model is at capacity. Please try a different model. Weird! But also good that it is being worked hard."

China. @akazwz_ on September 5, two days after the Astra launch (translated): "GPT-6 Astra load is a bit high, getting the error: Selected model is at capacity. Please try a different model."

LINUX DO, September 7 (translated from Chinese), in the thread about Astra:

"I've read a ton of posts and looked on X. Some people are completely unaffected, some have exactly what you have (me too). The most common theory is account risk control."

"Tested it. Doesn't really work. Even asking in pure English, Astra won't reply. 5.6 Sol is fine."

"Switch to a US node. The backend service is rate limiting. It's nighttime in the US right now, off-peak, so it won't throttle."

"I switched to a US node and it's working now, but the US node I bought is a bit slow."

LINUX DO, September 8, in a thread titled "[Codex] Selected model is at capacity. Please try a different model. Solution" (translated):

"It's been a whole day since yesterday and it's still like this. I'm about to lose it. What is the actual problem?"

"None of them work. Quota at 100%."

"I hit this last week going through a relay. Switched to logging in directly and it was fine. Apparently you need a good proxy for it to work."

And on the OpenAI forum, one Pro user called the outage "unacceptable" and asked for compensation. On GitHub, another wrote: "Top-tier service... if you're looking for a masterclass in how to ignore paying customers."

The pattern across all of it: paid users, plenty of usage left, error hits mid-task, and no fix from OpenAI beyond "try a different model."

How to fix "Selected model is at capacity. Please try a different model."

None of these are guaranteed, because the root cause is on OpenAI's side. But this is the order that gets people back to work fastest, based on what users report.

1. Do not reinstall, clear cache, or buy anything

Start here because it saves you 20 minutes. The error is not local. Reinstalling Codex, logging out and in, clearing browser cache, or topping up credits changes nothing. OpenAI's own contributor said the message is not account-specific.

2. Check the OpenAI status page

Open status.openai.com. If there is an active incident on ChatGPT or Codex, you are waiting on OpenAI and nothing below will help much. If the status page is green, the problem is regional or model-specific capacity and the workarounds below have a real shot.

3. Wait 5 to 15 minutes, then retry once

The single most reported fix. Capacity frees up in waves as other users' tasks finish. If you are in Codex, the app now retries automatically with a countdown, so you can let it run. But do not spam manual retries. Users report the backoff timer grows with each failed attempt.

4. In Codex, type "continue" instead of switching models

If the error interrupted an agent mid-task, you do not have to abandon the thread or the context. Sending a follow-up in the same thread like "Keep going without stopping" usually resumes the work once capacity is back. A Pro user in the July forum thread went further and built a small "alive keeper" bot that watches a Codex session and sends "continue" whenever the capacity error appears.

5. Switch to a sibling model, not a downgrade

Capacity is per model. When GPT-5.6 Sol is saturated, GPT-5.6 Terra or Luna often is not. When GPT-6 Astra is saturated, GPT-5.6 Sol is usually fine (the LINUX DO users above confirmed exactly this on September 7). Drop the reasoning effort from xhigh to high if that is an option, since the high-effort tiers are the first to fill. Going all the way down to GPT-5.5 works but most people describe it as a noticeable quality loss.

6. Start a new chat or session

Some users report that a fresh thread works when an old one keeps failing. Others report the exact opposite: their old session stays stable and every new one fails. It is cheap to test, so try both.

7. Try the CLI if the desktop app fails, or vice versa

During the June 16 incident, forum users reported Codex CLI working while the desktop app returned the error. On September 2 a Chinese team reported the reverse split: their coworkers on Codex's official client were blocked all morning while coworkers using third-party editors that talk to OpenAI through the Chat Completions API format kept working. Different clients hit different routes. If you have two ways to reach the model, try the other one.

8. Reduce concurrent sessions

Several reports mention running three or more Codex sessions at once, or sharing one Pro account across a team, right before the error appears. One LINUX DO user speculated that high concurrency triggers OpenAI's risk controls on the account rather than pure capacity. Unproven, but if you are running parallel agents on one account, drop to one and see if the error stops.

9. The VPS or VPN region switch (works for some, not guaranteed)

This is the fix people ask about most, so here is the honest version.

The error is about capacity in your serving region. OpenAI Support asks for your region or country when you report it. That means the same model can be full for you and free for someone routed through a different data center.

Some users report that spinning up a cheap VPS or VPN in another country and routing ChatGPT or Codex through it clears the error immediately. Singapore is the location that comes up most in reports we have seen, with users saying United States and Japan endpoints did not help them. The Chinese forum threads above tell a slightly different story: on September 7, one user cleared the error by switching to a US node during US nighttime, when American demand was low. The common thread is not the specific country. It is being routed to a region that is off-peak relative to where you are.

Caveats you should know before you try it:

• It does not work for everyone. In the July forum thread, a Pro user reproduced the error on a mobile hotspot with VPN and secure DNS turned off, so their problem was not network routing at all.
• A bad proxy makes it worse. Users on LINUX DO reported that going through a low-quality relay triggered the error, and logging in directly fixed it.
• OpenAI's official troubleshooting advice is to disable VPNs and proxies, not enable them. Routing through another country may also conflict with the terms of your plan depending on where you are. That is your call.
• Speed suffers. Users who made it work through a distant node complained the responses were slow.

If you want to try it: pick a VPS or VPN endpoint in a region that is currently in off-peak hours, connect, restart the ChatGPT or Codex app so it opens a fresh session, and retry. If it works, you will know within one message.

10. Shift your heavy work off US peak hours

Most of OpenAI's traffic comes from the United States, and the busiest window is roughly 7 AM to 5 PM Pacific. The April 23 wave of reports lined up with the start of the US morning. If you can run long agent tasks early morning or late evening US time, you will hit the error far less often. This is the boring fix that actually works.

Why OpenAI has not fixed this yet

Two reasons, and neither is going away this month.

First, demand. GPT-6 Astra is the biggest launch OpenAI has done, its own president said it may represent AGI, and every Pro subscriber tried it in the same 48 hours. GPT-5.6 Sol was already capacity-constrained in July. Adding a heavier model on top did not help.

Second, the failure mode. The most common request in every GitHub thread is the same: if a model is full, refuse new requests instead of killing running tasks. As of this week, Codex still cuts running work when capacity drops, then auto-retries. Until OpenAI changes that behavior, the error will keep interrupting agents mid-task even when overall capacity is fine.

What to do right now

If you are reading this because the error is on your screen: check the status page, wait ten minutes, then switch GPT-5.6 Sol to Terra or Astra to Sol and tell it to continue. That gets most people moving again.

If it keeps happening every day at the same time, you are hitting a peak-hour problem. Move heavy Codex runs off US business hours, keep one session per account, and if you are comfortable with it, test a VPS in an off-peak region.

And if you use Codex to run lead generation or scraping agents, build the retry into the workflow instead of babysitting it. We wrote up how we run Claude Code and Codex against Google Maps for local leads and the LocalProspects API it calls, which returns the data in one request so the agent is not sitting on a long model session when capacity drops.

Bookmark this page. We will update it as OpenAI posts new incidents or changes how Codex handles capacity.