“The latest version of Deepseek's V4 Flash model used 12% fewer tokens compared to its previous iteration.”