That's also not true at all. The most efficient languages for an LLM are the ones that have the largest datasets to reference. It knows the shape of a billion lines of Python. There's no evidence to suggest that "a more ergonomic language" burns less tokens cause it probably has to do multiple sweeps to get it "right". Otherwise some "agent/token efficient" language would've already been built and we'd all be using it. The defacto are just the most popular languages and that's not a coincidence.

Only if they gain that ergonomic advantage without complicating flow of control. Metaprogramming is costly for example, in the past so have been operator overloading and lazy evaluation.