The number of tokens in a model's reasoning or solution trajectory, often reduced through better training efficiency.