Rate limits, quotas and the Free allowance

Every limit answers in the same way: 429 with Retry-After for a limit on you, 503 with Retry-After when the server or the farm is busy. Your client handles both by waiting and retrying.

What can limit a request

AI Suite apps are never subject to the Free allowance. Pro Personal and Pro Commercial have no allowance at all; their limits are only the ones the operator sets.

What counts

Real work counts: chat, embeddings (a batch submission counts once), images, audio, vision and model downloads. Listing models, health probes, polling a batch job and monitoring do not.

How to respond

import random, time
import openai
def call_with_retry(fn, attempts=5):
    for attempt in range(attempts):
        try:
            return fn()
        except openai.RateLimitError as e:          # 429
            wait = float(e.response.headers.get("Retry-After", 1))
        except openai.APIStatusError as e:
            if e.status_code != 503: raise
            wait = float(e.response.headers.get("Retry-After", 2))
        time.sleep(wait + random.uniform(0, wait / 2) + attempt)
    raise RuntimeError("server still busy")

• Respect Retry-After; never retry in a tight loop.
• Mark unattended work with the header X-AISuite-Priority: background (or maintenance for the lowest priority), or run it as a batch job, so it waits while people are working. The header can only lower a request's priority, never raise it.
• Spread large jobs over time instead of firing them all at once.

Checking your usage

The operator sees usage per key, model and app on the server's Usage page or through GET /v1/server/usage with an admin key. Ask them for your key's limits.

Questions