Need to be careful with token counts. Looks to me by default "Automatic Result Analysis" is on and everything is sent to OpenAI leading to high token count.
Also, a tool's output can be Wolfram Language expression, for an image or diagram I think this again adds to the token count if one is not careful!
