A lot of production AI work is not conversation. It is deciding which of six categories a document belongs to, ten million times a month.

Why smaller often wins

• Latency you can put in an SLA rather than hope for.
• Cost that does not scale with a vendor’s pricing decisions.
• Data that never leaves your boundary, which removes an entire class of compliance conversation.
• Behaviour that does not change underneath you when a model is deprecated.

The last point is underrated. A hosted frontier model is a dependency whose behaviour can shift without a version bump on your side, and your evaluation set is the only thing that will tell you it happened.

Use the large model to build the training data and the small one to serve it. That combination has been the most reliable pattern I have seen.