Repository navigation
Codex service at capacity: widespread 'Selected model is at capacity' errors and hours-long unusability (multiple user reports, Sep 8 2026) #43738
Description
What issue are you seeing?
Summary
Since the Sep 8, 2026 DevDay / GPT-6 Astra launch, there is a large wave of user reports that Codex is effectively unusable due to capacity / server overload errors. The most common error:
Selected model is at capacity. Please try a different model.
Observed patterns across reports:
- Repeated "Selected model is at capacity" errors on any model selection, blocking normal use for hours
- Complete unusability for 5+ hours in some reports
- Paying Pro users (~$200/mo) unable to use the service at all; some requesting refunds
- Quota burn on Astra makes downtime worse (one user: a single ~2h Astra Medium task consumed ~80% of a Pro 5x weekly limit)
- Multiple reports that the capacity problems also affect the older Sol models
This appears service-side rather than client-side, and is not specific to one surface (CLI, IDE extension, and web users all report it).
User reports (X/Twitter)
Capacity errors:
- @AlexAndJust — repeats the capacity error, asks OpenAI to fix this first (20 likes)
- @xzfisher — begging for a fix for the capacity error (11 likes)
- @Kziii71 — same error, cannot do anything (4 likes)
- @mjwhhh — "selected model is capacity... please fix this" (2 likes)
- @Night_RevengeQ — same error (1 like)
- @Airsaid1024 — asks why this issue is not being responded to (1 like)
- @Dazui0225 — "so many capacity errors that I can't use Codex normally" (also reports overload)
Unusable / downtime:
- @avndp — "Codex is borderline unusable right now"; the most-agreed complaint in the thread (~55 likes)
- @BenoBilo61874 — capacity problems also affecting Sol, which was previously stable (5 likes)
- @eb_cryptodev — paid $200, cannot use it, requesting a refund
Usage-limit burn (compounding factor):
- @MaaSonder — top reply under the launch post (140+ likes), 5-hour limit frustration
- @SynKavish — one ~2h Astra Medium task burned ~80% of a Pro 5x weekly limit (6 likes)
- @iamMarketWiz — asking for Astra limits to match Sol (4 likes)
- @AskGodForWisdom — ~15 min of coding then ~4h45m blocked, repeating for days (3 likes)
- @Nishendigital — usage ran out after ~40 min of work (2 likes)
- @pierizvi — Plus quota "vanishes in a min" on Astra
Source: the replies under @thsottiaux's launch post (~400 replies, dominated by these complaints).
What steps can reproduce the bug?
- Since ~Sep 8, 2026 (post-DevDay), open Codex on any surface (CLI, IDE extension, or web).
- Start a task with any available model.
- Frequently fails with:
Selected model at capacity. Please try a different model. - Switching models and retrying often hits the same error; degraded windows last for hours.
Not a client bug — filing as a service capacity / observability report.
What is the expected behavior?
- Selected models should be usable during launch windows, or users should get a clear queue/status message instead of an error loop
- Quota should not be consumed by failed attempts caused by capacity errors
- Paying users should receive acknowledgment (status page / compensation) for extended downtime
Additional information
Report compiled from public X and GitHub reports. Happy to add more user reports as they come in.
Activity
github-actions commentedon Sep 8, 2026
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
linxuanweb commentedon Sep 8, 2026
same issue, any updates?
only 2% usage limit spent this day due to this shit bug.
it's really a p0 bug!!!
mohammadbintangbiscirurgiputra commentedon Sep 8, 2026
Thanks for confirming, @linxuanweb — adding your report to the pile. Same here: quota barely touched because runs keep dying to "Selected model is at capacity".
Agreed, this is P0. Two things that help:
It is not fair that paid users are just watching their time get wasted while this goes unacknowledged.
YannCedric commentedon Sep 8, 2026
Same issue.
As a pro 20x user, this is super frustrating, it's been a few days I can't use any models in codex (v0.153.4)
zhuylanz commentedon Sep 10, 2026
I'm experiencing the same issue. As a Pro 20x User, I'm supposed to have prioritized access, but it keeps saying this and failing without any progress. This is unacceptable.
KSimulation commentedon Sep 10, 2026
LOL, this is a FEATURE, not a BUG.
wenner commentedon Sep 11, 2026
September 9th was fine, but the 10th was completely unusable... and it's been like this ever since. This includes 20x, 5x, and Plus, all experiencing similar issues, but 20x is the most severe, rendering it completely unusable.
That's outrageous! There's no after-sales service, which is completely unacceptable.
cartesius-3 commentedon Sep 15, 2026
今天是9月15日这两天还是会碰到这种情况,不过比前几天情况好太多了,还是没有完全解决吗
cartesius-3 commentedon Sep 15, 2026
并且我做了对照试验,同一个20xpro订阅账号,在windows设备上频繁出现,mac上却没有问题
ZhmAAA commentedon Oct 8, 2026
已经持续两天了，我今天还是这个问题 Mac
away0730 commentedon Oct 12, 2026
Still down as of October 12, 2026 — GPT-6 Astra and GPT-6.1 Sol both unusable
Adding another data point to this thread. As of today (October 12, 2026), I am still experiencing the same capacity errors described in the original report.
Both GPT-6 Astra and the newer GPT-6.1 Sol are affected for me. I keep hitting “Selected model is at capacity” across the CLI and the web version. Switching between models doesn’t help — the error follows regardless of which one I pick, which matches what others have already noted in this thread.
It’s been over a month since the original reports on Sep 8, and the issue appears to be ongoing. Is there any official update on when this will be resolved?
Aether-254 commentedon Oct 19, 2026
双Plus 6.1sol/5.6sol/6astra/6sol均此问题
YanArty commentedon Oct 19, 2026
same sht, I plan to switch to Claude code x20
gpt 6 astra, gpt 6.1 sol worst models ever in my practice, same sht with gpt 5.6, all are depredated status(
ruined 3 projects same time using gpt astra and new sol 6.1, not recommend, 10 bug reports, seems openAI done( sad
Artypoul commentedon Oct 20, 2026
It seems like this is a kind of ban.