You don't need SOTA models for all tasks, and being able to do more routine tasks at something like 10% of the cost and 70x speed unlocks LLM use for things that are just unthinkable now (bulk classification tasks, real time speech interaction, etc)