A fact many people don't know: Conversing with AI in Korean is 50%–70% more expensive. In English, roughly one word corresponds to one token. "Hello" is 1 token, and "artificial intelligence" is 2 tokens. However, Korean is a bit different. "안녕하세요" (Hello) is broken down into 2 or 3 tokens. Due to its combinatorial structure, Hangul is structurally designed to use more tokens than English. Writing the same content in Korean consumes about 1.52 times more tokens than writing in English. Since API costs are proportional to tokens, Korean is a whopping 50%–70% more expensive for the same content. Furthermore, if you receive the AI's response in Korean as well, the output tokens are likewise 50–70% more expensive. And regarding this, some people have mentioned that there are other research results (arxiv.org/pdf/2507.00246), This is a study that only tested mathematics, and it even completely excludes models we frequently use, such as GPT and Claude. These models receive RLHF based on English, so the results may differ. The models used in this study are DeepSeek R1, Qwen 2.5, and Qwen 3, so they are all LLMs originating from China… Also, the premise that "token reduction = efficiency" is a bit problematic. Even if thinking in Korean reduces the number of tokens, the cost of one Korean token is higher than one English token (byte count, processing cost). So, in conclusion, if you use AI frequently and keep reaching the rate limit, I recommend that you just converse in English, treating it as a form of English practice for now haha.