New tokenizer has a much larger vocabulary (200k)[0].

[0] https://github.com/openai/tiktoken/commit/9d01e5670ff50eb74c...