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