On OpenAI's current o200k_base tokenizer, the same sentences need 7% more tokens in Chinese, 18% more in German and 54% more in Hindi than in English.
On the older cl100k_base tokenizer, Hindi needed 4.36 times the English token count, so the old penalty was far larger than the new one.
Chinese is the cheapest non-Latin script on o200k_base, which contradicts the popular claim that Chinese and Japanese cost two to four times more.