"Outperforms" is a stretch; its skill is on par or slightly exceeds the state-of-the-art NWP models currently run today, but you as an end user would not notice the difference between GraphCast and the ECMWF HRES. Furthermore, raw model guidance is almost always passed through a blending and statistical post-processing process alongside competitor models to fine tune the forecast, and there are no rigorous benchmarks of whether or not the inclusion of GraphCast improves this process because very few groups are running it operationally.

Also, it's worth noting that GraphCast's outputs are a tiny subset of what we would traditionally forecast as weather parameters. You can't out-perform a competitor on a task that you aren't solving!