I have a product that writes AI-written nonfiction books.
I expected more thinking to help, but Sonnet 5 on high effort wrote the weakest prose of any setting (scored by my own blind reads + different LLM judge), and it was slower and more expensive too with some chapters running out of tokens mid-sentence.
I agree that the voice is hard to edit out afterward - in my tests, which model & settings I picked changed it more than prompting/instruction.
