How do you choose between frontier vs. open-weight models for a new product feature?

We keep revisiting this for each new feature: GPT-4o / Claude for quality and speed-to-ship vs. Llama / Mistral for cost control and data privacy. The right answer seems to depend on so many variables — task complexity, data sensitivity, latency budget, team capacity to maintain fine-tuned weights. What framework do practitioners actually use to make this call, and when do you revisit it?