Didn't expect Claude to actually have a 'Chinese tax.' AI researcher Aran tested various models by having them translate the same content into different languages. Turns out, when translating to Chinese, Claude uses 65% more tokens. Other models, like Gemini, Qwen, DeepSeek, and Kimi, are much more reasonable in their token usage. Even OpenAI's ChatGPT only uses 15% more. The good news is, Claude also 'taxes' other languages. Among them, Korean gets hit the hardest, requiring a whopping 159% more tokens.