I had three frontier models (Opus 5.5, Sol 6, Kimi K3) produce two landing pages for a fake startup: one from an adjective set, one from a named reference. I then blind-rated 90 of these pairs.

I chose the named reference 68% of the time! Sol 73%, Kimi 67%, Opus 63%. Examples: https://t.co/8r1KfD07oJ