I see! So for these, you tend to find the accuracy "good enough" on the faster-but-less-accurate models.

I generally find the same thing for simple definitions/translations and other "chat" tasks. I'm a little bit surprised that you also find it so for coding, but otherwise I think I get it.