The interesting question is rarely which model. It is what the thing is allowed to do, what it does when it is unsure, and what it costs per month once real people are using it.
Not unless you ask for it. Standard builds use retrieval against your own data with commercial API terms that exclude training. This is written into the contract rather than assumed.
You get a per-month estimate based on your real expected usage before the build starts. Running cost is a design constraint from the beginning, not a surprise afterwards.
Anything customer-facing is designed to cite its source and to say it does not know rather than guess. Where a wrong answer would be expensive, a human approves before it goes out.