The author recently noticed that after the release of ChatGPT 5.2, the thinking time in thinking mode seems to have shortened. To verify whether the model performance has declined, the author conducted a juice value test. In extended thinking mode, it was observed that the model sometimes outputs 256 tokens, but sometimes cannot provide the same output, showing unstable performance. The author posted on the Linux.do forum to ask if other users have encountered similar situations, and 6 posts have participated in the discussion, involving 5 participants. Hope to collect more data to analyze model performance changes and provide reference for users.
