"They" don't "care" about anything. It is a stateless computational run across thousands of semiconductors. There is no objective this software has other than the computational function completing. To care would mean the model would have a level of discernment that goes along with sentience.

> 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?

You clearly haven't used Claude to generate code or documentation.