I assume they will have a lot less "safety", i.e. the model will be more likely to actually do what you ask instead of finding a reason why "sorry Dave, I can't do that".

Since these "safety" features tend to also degrade the model, that's likely also helping them catch up in the benchmarks.