It really seems too lucrative for them to weight the models toward higher token usage and less-deterministic solutions for them not to do it.