If one provider's outage takes your feature down, you need a fallback path, and you have three realistic options. A hosted router such as OpenRouter is the least code but adds a third party to your request path. A self-hosted layer such as LiteLLM keeps keys and logs in your hands but is something you now operate. A 40-line wrapper around two SDKs is enough when you only ever call two models.
The symptom that sends people searching for this looks like 529 overloaded_error , 503 Service Unavailable , or a stream that opens and then dies halfway. I'll walk through what to fix first, then compare the three approaches. Model names below are placeholders; substitute whatever you run as of November 2026.
Why does retrying the same provider not fix a 529 or 503?
Retries handle blips. Overload and regional incidents last minutes to hours, and a retry loop with backoff just queues your users behind the problem.…