The benchmarks are very long form logic, knowledge, and coding tasks though. I'm very interested in Haiku 5.5's performance on ObviousBench where Luna 6 is currently SotA.