If you are looking at huge categorizing and parsing work, using a local decision model and passing it to Haiku or Luna can really lower your token cost.