I reduced image-processing token usage by ~95% compared with GPT-4o direct vision, while maintaining roughly the same accuracy.How significant is that?[P]