You fall back to a model you've never evaluated with your prompt. What can go wrong?

It's a different model: format reliability, refusals and quality may all change, turning an outage into a quality incident. Give the fallback path its own prompt version if needed, its own evaluation run, and metrics split by which model answered.