One thing I didn't expect: the models differ a lot in how "faithful" they get when you type a real URL. Of course smaller models are further from the truth, but it seems to me that the big models from OpenAI and Anthropic tend to try to prevent copying a branded web site, while the Chinese model GLM 5.2 often delivers almost 1:1 copies of the layout of sites like wikipedia.org etc. Feels like a training/RLHF difference around reproducing recognizable content. Does anyone have more insight into that?