TL;DR * Reliable AI in production depends less on model quality than on what happens when a provider call fails, because 429 rate limits and 5xx errors arrive on a routine basis. * Retries and fallbacks solve different problems: retries handle transient errors inside one provider, fallbacks switch providers once the

TL;DR * Reliable AI in production depends less on model quality than on what happens when a provider call fails, because 429 rate limits and 5xx errors arrive on a routine basis. * Retries and fallbacks solve different problems: retries handle transient errors inside one provider, fallbacks switch providers once the