“The same document can have a 10-20% variation in token count between different model providers, with the variance being even higher for Anthropic's Claude and non-English text.”