GPT-3 (ChatGPT) also succeeds at deciphering your example text. I didn't think it was that impressive until I realized the tokens were going across word boundaries like you said.
GPT-3 was silently replaced with RLHF'd ("instruct") versions (while keeping the same model names in the Playground) because they were cheaper to run (and had better "safety").
I think this means that current GPT-3 is more similar to GPT-3.5-turbo than to old (retired) GPT-3, but I'm not entirely sure about the details.
ChatGPT's lower model is GPT-3.5-turbo, it is not GPT-3.