I belive after the next generation of Claude/OpenAI models the frontier is gonna shift to inference speed.
Any increase in intelligence is marginal at this point as 90% of the tasks are one-shotted or require few iterations.
The biggest bottleneck now is speed, opus5.5 can do great work but it takes multiple days for complex tasks increasing the cost of misalignement on the initial promps.
Improving alignment is great and all but if delivery time would be 5x faster the better iteration speed would be even more valuable