When would you choose a less capable AI model?

From a purely technical perspective, what’s the case for choosing a less capable AI model when you have access to a stronger one?

Set cost aside for a moment. I’m curious about the engineering reasons: latency, privacy, deployment constraints, predictability, or better performance on a specific task. And how do you decide whether those advantages outweigh the difference in capability?

What’s a concrete example where you’d deliberately choose the less capable model and the why behind it?

#AI #LLMs #AIEngineering