Absolutely agree. Granted, it is task dependent. But when it comes to classification and attribute extraction, I've been using 2.0 Flash with huge access across massive datasets. It would not be even viable cost wise with other models.