Codex "Selected model is at capacity": is it you or OpenAI?
It is OpenAI running short of capacity for that model, not your account or usage limit. Measured on real Codex requests through the Wokey gateway, 1.3% of requests hit this error in the last 24 hours (status: Outage). Since 2026-05-31, 5 outages pushed the daily rate above 2.0%; the worst day, 2026-09-10, reached 29%.
Updated 2026-10-12 02:08 UTC, every hour
Right now: Outage
- Last hour: 13% (254 requests)
- Last 24 hours: 1.3% (18,655 requests)
- Last 7 days: 1.7% (198,720 requests)
Status reads the last three hours: under 0.5% of requests is normal, 3.0% or more is an outage. OpenAI status page
Error rate by model: which one to switch to
Capacity is per model, so switching models is the one fix that works. Lowest 7-day rate first; models with fewer than 1,000 requests in 7 days are not listed.
| Model | Last 24 hours | Last 7 days | Requests (7 days) |
|---|---|---|---|
| GPT-5.6 Sol | 0.7% | 0.8% | 58,984 |
| GPT-5.6 Luna | 1.9% | 0.9% | 54,961 |
| GPT-6 Sol | 0.2% | 1.9% | 18,460 |
| GPT-6.1 Sol | 1.3% | 2.4% | 53,435 |
| GPT-5.6 Terra (Most stable now) | 0.0% | 5.0% | 2,737 |
| GPT-6 Luna | 2.3% | 5.1% | 6,496 |
| GPT-6 Astra | 3.6% | 5.3% | 3,094 |
To switch in Codex, type /model and pick gpt-5.6-terra, or set model = "gpt-5.6-terra" in ~/.codex/config.toml.
Outage history
Daily share of Codex requests that hit the error. A day at 2.0% or more counts as an outage day; outages one quiet day apart are merged.
- 2026-10-08 – 10-12: 5 days, worst on 10-09 at 5.4%
- 2026-09-21 – 09-24: 4 days, worst on 09-21 at 4.9%
- 2026-09-14 – 09-17: 4 days, worst on 09-17 at 28%
- 2026-09-06 – 09-10: 5 days, worst on 09-10 at 29%
- 2026-08-19: 1 day, worst on 08-19 at 5.3%
What to do
- Do not buy credits, reinstall Codex or delete the task: none of them changes OpenAI capacity.
- Check the status above and the OpenAI status page. If the rate is elevated, it is not you.
- Retry once after a minute; most single failures pass on the next try outside an outage.
- During an outage, switch to the most stable model in the table instead of retrying the same one.
- If a task stopped halfway, send "continue" in the same thread once it works again.
- Only if your account alone keeps failing for days, open a support ticket or GitHub issue with the session ID.
Is it my account?
Almost never. In the 2026-10-08 – 10-10 outage, 91% of failed requests had another account succeed on the same model within two minutes, and 41% had the same account succeed within two minutes. 57 / 74 accounts were hit, none of them more than others. OpenAI rejects a share of requests at random while a model is short of capacity; it does not single out accounts.
"At capacity" is not a usage limit
A usage limit says your plan is spent until a reset time. "Selected model is at capacity" says the model is short of servers right now; your remaining usage is untouched and the error clears without waiting for a reset.
How we measure
- Source: real Codex requests through the Wokey gateway, counted every hour by model. No user data is published.
- A capacity error is a request the upstream rejected with server_is_overloaded ("Our servers are currently overloaded"), which Codex shows as "Selected model is at capacity". Wokey does not retry these, so every one reached the client.
- Only finished requests count. Image models and ChatGPT on the web are not included.
- Our traffic mix is not the same as all Codex users, so read the rates as a measured signal, not as OpenAI's official figure.
FAQ
Why does Codex say "Selected model is at capacity"?
OpenAI is short of servers for that model and rejects part of the requests with server_is_overloaded. It usually happens in waves after launches, such as GPT-6 reaching all ChatGPT users on 2026-10-07.
Is it my account or my usage limit?
No. Your remaining usage is not touched. In the 2026-10-08 – 10-10 outage, 91% of failed requests had another account succeed on the same model within two minutes.
Which model should I switch to?
Right now GPT-5.6 Terra has the lowest measured error rate (0.0%). The table above is updated every hour.
How long does it usually last?
Since 2026-05-31 the outages we measured lasted 4 days at the median, the longest 5 days. Within an outage the rate swings hour to hour.
Is "Our servers are currently overloaded" the same error?
Yes. It is the upstream message behind the same server_is_overloaded code; Codex shows it as "Selected model is at capacity. Please try a different model."
Does buying credits or upgrading to Pro help?
Not for this error. It is about the model's capacity, not your plan. Reports on openai/codex include Pro users hitting it with most of their usage left.