What impressed me most about Claude Sonnet 4.5 is that its output structure is more stable than many other models and less prone to crashes. I ran some real world scripts from my own projects, and it exhibited fewer hallucinations than GPT-4 and performed more faithfully on code interpretation tasks. However, it can be a bit slow to warm up, and sometimes I needed more prompts in the first few rounds.