I ran it today on 200 descriptions for the sim racing shop, one product per call, nothing fancy.
You're right on the tokenizer. Same text, more tokens. I got around 12 percent more than the old one in the usage numbers.
Still cheap though. For a short product text the whole batch of 200 was under one cent in input and that's the case where the 100K jump never comes close because nobody puts a short description in a prompt that long.
I'm using it already for that job. For the long catalog ones I stay on the old setup.