Five silent ways LLM cost tracking tools produce inaccurate spending data

A developer building an LLM observability service discovered five distinct ways that self-built cost metering layers silently produce wrong numbers, with no errors or alerts to signal the problem. One major issue is that OpenAI's Chat Completions API returns zero usage data for streaming requests unless a specific option is explicitly passed, causing most chat traffic to be recorded as costing nothing. Prompt caching further distorts figures differently across providers: OpenAI's cached tokens lead to over-counting, while Anthropic's separate cache fields cause under-counting if ignored. The author warns that these failures are especially dangerous because they can silently disable budget-gate features, allowing unchecked spending to continue undetected. The findings are presented as a practical checklist for teams building their own LLM usage tracking infrastructure.