I used @OpenRouter to generate texts from 18 different LLMs across a range of genres. Then scored them with @Havelock_AI Some findings: Grok the most oral gpt-4o the least (surprising) Chinese labs generally default more oral than US ones, particularly in certain genres. (Joe Weisenthal)