Production checklist

Keys & caps

• Use one key per environment (dev, staging, production) so that one surface can be revoked without disrupting the others.
• Rotate by creating a new key, switching over, then revoking the old one. Revocation is immediate.
• Set a per-key monthly spend cap to bound blast radius. Past the cap a key returns 402.

Handle errors & retries

Retry 429 and transient 5xx responses with exponential backoff, honoring the Retry-After header. Do not retry 401 or 402: fix the key or top up first.

import time
from openai import OpenAI, RateLimitError

def call_with_retry(client, **kwargs):
    for attempt in range(5):
        try:
            return client.chat.completions.create(**kwargs)
        except RateLimitError as e:
            wait = int(e.response.headers.get("retry-after", 2 ** attempt))
            time.sleep(wait)
    raise RuntimeError("exhausted retries")

See Errors and Rate limits for the full envelope.

Monitor spend

Poll GET /key to read remaining budget and usage, and keep the wallet funded so calls are not rejected at 402. The console usage page shows per-call cost.