the most impressive new thing i've seen from gpt-5.4 is that it seems to be trained to abandon unpromising directions instead of hitting the head against the wall. this is a great improvement in agency and intelligence, sometimes things don't have a clear right answer like in math problems the models are trained on.