I remember that in the past (just a few hours ago at the time of writing this post), before the ChatGPT update to GPT-4 Turbo, there was a token limit of what I could write in the chat. It was something around 3080 tokens or something close to this number.
Now I was curious to see what the new token limit might be. So I looked at the tokenizer and found that ‘xj3’ is three tokens and I can repeat this sequence and keep adding tokens. So ‘xj3xj3’ would be 6 tokens. And every single char is one token. Good.
Now I did this on my local machine (Python):
def filli(n):
os.system(f'echo -n {"xj3" * n} | xclip -selection clipboard')
On Ubuntu you might first have to apt install xclip.
After some playing around I came to filli(43676). Pasting this into ChatGPT worked. But increasing this number by just 1 would give me the info that my input was too long.
So with 43676 the prompt contains 131028 chars (and tokens). I have added another ‘x’ and it still worked. But now, adding the next ‘j’ resulted in an aborted request.
So my current best guess is that 131029 is the “real” new token limit. It might be the same for the API, which I haven’t tested yet (i.e. over one Euro per test).
This token count brings us very close to 2^17 = 131072
So the system somehow needs the remaining 43 tokens for something internal (it would have been fun if this were exactly 42 though 
).
Hope this helps.