Their voice capable model is several generations behind the state of the art text-only one, as far as I know.

I don’t think it even has reasoning tokens, so it’s no surprise that it’s as most as smart as the “instant” models (i.e., not very).