LLM APIs limit requests per minute, tokens per minute, or concurrent streams. Hitting the cap returns 429s; the client should back off, not retry immediately. Product rate limits and model-provider rate limits stack, which is why an app can fail while the lab dashboard still looks fine.