I'm amazed they didn't test xhigh thinking mode explicitly to ensure it didn't exceed the 128k thinking budget allocation. I guess pace of development gets away from everyone, even OpenAI.