FWIW Seeing the same thing with Opus 5. Not sure if the models are being over optimized for agentic tasks or over reliance on synthetic data, because since Opus 4.6 I feel like emotional/conversational intelligence has been on a decline (Fable being the exception)