Three Sume throttle signals: rate_limited, queue_full, 503

OpenAI's changelog for September 2, 2026 says API errors now distinguish traffic increasing "too quickly" (429) from temporary model overload (503). Sume has three throttle-like signals: 429 rate_limited, 429 queue_full and 503 provider_capacity_exceeded. Each needs a different response from your client.

The OpenAI change

From the OpenAI API changelog, read on 2026-10-03: a 429 now means traffic is increasing too quickly, and a 503 means temporary model overload. Sume's codes are its own and are described below.

What each Sume signal means

The error body carries a code and a request id. Concurrency being full is not an error by itself: Sume accepts valid jobs as queued while queue capacity remains.

Why queue_full is not a rate limit

queue_full means Sume cannot accept another paid job for the workspace until a queued or processing job finishes or is canceled. Sending requests more slowly does not fix it; finishing work does. rate_limited is about request volume, including polling.

One retry function

Never retry an unsafe submit without an Idempotency-Key, and reuse the same key across attempts so a retry returns the original job.

import time
import requests

def submit(url, headers, body, key, tries=5):
    headers = {**headers, "Idempotency-Key": key}
    for attempt in range(tries):
        r = requests.post(url, headers=headers, json=body, timeout=30)
        if r.status_code < 400:
            return r.json()
        code = r.json().get("error", {}).get("code")
        retry_after = float(r.headers.get("retry-after", 2 ** attempt))
        if code == "queue_full":
            retry_after = max(retry_after, 30)
        elif code not in ("rate_limited", "provider_capacity_exceeded"):
            r.raise_for_status()
        time.sleep(retry_after)
    raise RuntimeError("still throttled")

Log the request id

Record error.request_id with every throttled response. It is safe to share with Sume support. Do not log API keys or signed URLs.