I’ve heard what is now considered high accuracy is the 3-5 day ECMWF.  

I’ve heard that gets pushed out roughly 1 day per decade due to better modeling (maybe Snodgrass?).  I assume that is mostly raw computational power (higher resolutions, more runs), with a side of better initialization data and probably some tweaked modeling/algorithms.  Not sure how that squares with Moore’s law, but it definitely seems to be getting slowly better.