Far better results compared to GPT 5.4 and Opus 4.6. Not great for execution due to speed but has consistently had better comprehension of the codebase. Maybe it's a case of "holding it wrong" regarding the other models but that's been my experience.