Token counts from the OpenAI tokenizer page don't match what the API bills

Three things stack up. First, encodings differ per model, the web page defaults to one and your script may be using another, use tiktoken.encoding_for_model with the exact model name. Second, the API counts the chat framing, every message costs a few tokens for role and separators plus a couple for the reply priming, so a one line prompt bills noticeably more than its text. Third, system prompts and any tool definitions are counted too. Trust tiktoken with the right encoding for the text, add the overhead, and treat the usage block in the response as the real number.