Best Practices for Model Selection: Is 'Auto' mode reliable for complex full-stack tasks?

Background:
I am a full-stack engineer working on both frontend and backend simultaneously.

My Current Workflow:
I’ve been sticking to the most powerful models (like Opus 4.5) for almost everything. The completion quality is excellent, and it handles complex context very well. However, it can occasionally be slow.

The Question:
I’m intrigued by the “Auto” model selection , which promises a balance between speed and quality. However, I have concerns:

  1. Trust Issue: Does “Auto” successfully identify when a task is complex? I’m worried it might route a tricky logic problem to a smaller model, resulting in buggy code or hallucinations.

  2. Quality Drop: For those who use “Auto” or smaller models regularly, do you notice a significant drop in code quality compared to strictly using Opus 4.5?

What I’m looking for:
I’d love to hear your experiences/strategies on model selection. Do you leave it on “Auto”? Or do you manually switch depending on the task?

Thanks for sharing!

See also:

It does a good enough job for full stack development for me but is a bit on the slow side.

If that doesn’t bother you Auto is completely fine unless you’re doing anything novel

By it I mean whatever model is currently being used in Auto mode as it’s a bit of all sorts of various cheap models