depends for what.

For autocompleting simple functions (string manipulation, function definitions, etc), the quality bar is pretty easy to hit, and speed is important.

If you're just vibe coding, then yeah, you want quality. But if you know what you're doing, I find having a dumber fast model is often nicer than a slow smart model that you still need to correct a bit, because it's easier to stay in flow state.

With the slow reasoning models, the workflow is more like working with another engineer, where you have to review their code in a PR