Comment by p-e-w

That's true, but not my point. My point is that if the request specifies GPT-3.5, Nitro knows that it cannot possibly serve that model, so anything other than returning an error is simply lying to the client, which is a really bad idea.