622 of our 5,087 LLM API calls never returned an answer

Every tutorial counts calls. Your bill counts attempts. Here is the real error distribution from 5,087 production calls, which failures are worth retrying, and why a timeout is the only one that actually burns tokens.