Do you pick Claude/GPT before or after writing the task test?

Model choice framed by acceptance criteria rather than brand preference.

I have started trusting model comparisons only after the task has a pass/fail shape. For code work, that means diff quality and test behavior. For research, it means source handling and correction behavior. Without that, Claude vs GPT becomes a vibe check. Useful, but not enough to choose a team default. What would you write as the smallest acceptance test before choosing the model?