The cost of training/developing a frontier model will continue to rise. The cost of inference will continue to fall. Companies don't buy models as such, they buy inference. As inference becomes a larger and larger business expense companies will optimize that spending, routing work to the "capable enough" least cost tokens.

There is a lot of human work that can be offset with tokens. On the other hand, the amount of work for the highest intelligence models will remain limited - corporate strategy, major decisions, etc. are just not as frequent.