The models benchmarked by RULER do worse in needle in a needlestack. It will be interested to see how 4o does with RULER.