Just tested it myself—under GPT Work mode, a short conversation only uses about 200~300 tokens. Using Skill to refine a 7,400-word article (I provided my own draft), it ended up consuming a total of 25,000 tokens. So scary