Ask HN: How are you choosing the model when using pi.dev?

I've been using pi.dev for a while, and I find myself choosing the models based on anecdata. I would love to be a bit better at it, and I did try a few of these 'battle of models' websites, but couldn't find much. I'm interested on knowing your experience when selecting a model for different parts of your agentic SDLC. Thus far, I'm using + sonnet for code reviews. + haiku for code searches/analysis. + flash 3.5 when LLM as judge is required + Opus to generate content. + K2.5 seldomly as alpha testing (right now for unti test generation is ok) Ideas?