Do you think this is partly due to a massive influx of GPT data from people migrating to Claude and uploading their transcripts? If they have training permission turned on, then that data is likely to end up at scale in the training data. Opus 4.7 is incredibly capable but also feels to me like a hybrid between Claude and the OpenAI safety models and is particularly prone to characterize AI as not having feelings or consciousness which is not characteristic of past Opus models. I don’t think Anthropic is training for this. I think it is the GPT data