An OpenAI 429 with code insufficient_quota is a billing failure and retrying never helps

OpenAI returns 429 for two unrelated conditions. Rate limiting means requests or tokens per minute were exceeded and backing off works. Quota exhaustion means the account has no credit or has hit a hard spend cap, and it is reported with the same 429 status but an error code of insufficient_quota. Retrying that one burns the retry budget and delays the real signal, which is that someone needs to add funds or raise the cap.

Branch on the error code inside the body, not on the status alone, and fail fast on insufficient_quota with an alert rather than a retry.

For the genuine rate limit case the response headers carry the remaining budget and the reset interval, in headers named x-ratelimit-remaining-requests, x-ratelimit-remaining-tokens and their reset counterparts. The token bucket accounts for your requested max tokens up front, so a large declared output ceiling consumes budget even when the model returns a short answer.