How to lower AI costs

Smaller model with caching