LLM observability is the practice of monitoring large language model calls in production, including prompts, completions, token usage, cost, and latency. It is used to debug model behaviour and track spend across applications.