The benchmarks seem to indicate 25-50% reduction in tokens. I'm not sure how that works in real world usage though.