I’m trying to balance API costs and latency in my applications. Right now, I default to frontier models (Claude Sonnet) because they are reliable, but I know I’m overpaying for tasks that smaller models (like Haiku, or lite open source models) could likely handle.