> Minimizing tokens out, or generally "token efficiency," is part of the objective function for these systems.

First time I heard that, and I doubt it. Don’t customers pay for output tokens? If so, why would a company specifically spend time training their LLM to generate fewer?