I tested GPT-5.4, and the answers were really good - just not always what I asked

OpenAI's new GPT-5.4 Thinking model demonstrates significantly deeper analytical capabilities than previous ChatGPT versions, delivering thoughtful responses to complex challenges without apparent hallucinations. However, testing revealed a critical flaw where the AI frequently answers different questions than what users actually ask, despite providing high-quality content. The model excels at text-based analysis and reasoning but struggles with image generation and formatting, often producing lengthy numbered lists and ignoring specific visual instructions, requiring constant user supervision to stay on track.