Haven't read the paper, but they are probably using something like sentencepiece with sub-word splitting and then charge by the number of resulting token. http

Haven't read the paper, but they are probably using something like sentencepiece with sub-word splitting and then charge by the number of resulting token. https://github.com/google/sentencepiece https://github.com/google/sentencepiece