Comment by acgourley

I disagree, their findings should generalize to the frontier. Even if the latest can deal with the extra complexity, it stands to reason it will take more tokens to do less. This could be a useful insight into the next generation of evals.