Maybe it would benefit Anthropic if AI generated code worked, but wasn't readable by humans. That's a nice moat.

Good variable names are still useful for LLMs to understand context when refactoring.

Only if LLMs will start to output object code, skipping text representation.