At this stage, GPT 5.4 + subagent actually doesn't need much further improvement in model intelligence.

Medium thinking capability = a standard programmer you'd recruit off the street.

If you use some standard processes + organizational structure, you can make them perform at 1+1>2 capacity.

Whereas a person who blindly relies only on xhigh thinking capability might only perform at 0.5 capacity, and it's also extremely slow.