gemini models solve a problem in 80% less tokens so that's something to think about.