server_is_overloaded on gpt-6-astra: error is account-specific and doesn't resolve on its own

⚡ The gist of the dispute

A thread by RE-217 is open on OpenAI Community: the error "Selected model is at capacity" with the code server_is_overloaded keeps recurring on the gpt-6-astra model. The author raises two complaints at once — the failure is tied to a specific account and doesn't resolve on its own. The thread has gathered 18 comments and 1231 views with zero votes.

🗣 What participants are saying

One caveat right away, without which this piece would be a lie: the texts of the 18 comments were not provided to the editorial team. So below is an analysis of what is claimed in the title and the problem statement — no quotes, no usernames, and no attributing positions to anyone. Who exactly replied in the thread and what they said cannot be reconstructed from the materials provided.

The title contains three separate claims, and they are verified differently. "Repeated" is about reproducibility: the error keeps coming back again and again. "Account-specific" is about being tied to an account rather than to the model as a whole. "Never auto-recovers" is about the absence of self-recovery. If the third claim is true, retries and backoff alone won't solve the problem — this is no longer a transport-level failure.

Then an internal contradiction begins, around which threads like this are usually built. The code server_is_overloaded is server-side by name: it speaks of overload on the service's side. But the behavior described by the author is account-level: other users of the model, by the logic of the complaint, are working fine at the same time. Either the overload is selective and affects only some routes, or the issue lies in the configuration and limits of a specific account — in which case the word "capacity" in the error text is misleading.

The platform's metrics reveal one more detail: 1231 views against 18 comments and zero votes. It looks like a thread that is being read en masse by people hitting the same error, but no settled answer worth upvoting has emerged in it. This is an observation based on numbers, not the participants' opinions.

Separately, it's worth keeping in mind what is entirely absent from the materials provided: no support response, no link to a status page, not a single attached log with a request-id. Without them, any analysis remains at the level of hypotheses.

⚖️ Where they agree, where they diverge

There is no common denominator in the materials provided — and presenting something as one would be a substitution. The participants' positions were not provided, so claiming "the community believes" is impossible: there is only the author's problem statement and a comment counter.

The real point of conflict lies in the diagnosis, not the treatment. An error with a server-side code is behaving like an account-level one, and until these two versions are separated by an experiment, the dispute is about whose failure it is: the infrastructure's or the request configuration's. The practical takeaway from the thread is one — without a request-id and a comparison of two accounts, the conversation won't move forward.

🧪 What to check yourself

1. Run the same prompt from two accounts or projects with a single key — this tests the "account-specific" hypothesis.
2. Compare gpt-6-astra with neighboring available models in one script: if only one fails, the issue is in routing to the model, not in the account.
3. Log the entire response — HTTP status, retry-after, request-id, org/project id, timestamp. Without a request-id, support has nothing to match against.
4. Run backoff with jitter for 10–20 minutes: if the error code doesn't change and no retry-after arrives, "never auto-recovers" is confirmed on your side already.
5. Check the status page and your tier/limits at the moment of the failure: a discrepancy of "the page is green but the error is persistent" is the strongest argument in a ticket.

🔗 Sources

OpenAI Community — Repeated "Selected model is at capacity" (server_is_overloaded) on gpt-6-astra — account-specific, never auto-recovers

Disclaimer: this is a summary of a discussion, not facts verified by the editorial team. The opinions belong to the participants; the numbers belong to those who provided them.