After 10 text and 4 image tests, OpenAI's latest model barely beats GPT-5.1. What are Plus subscribers really getting?