I guess the only explanation is that there's a side-telemetry channel that still sends some data to Anthropic, regardless of ANTHROPIC_BASE_URL overrides.
I'm quite all right with the first, not with the second of course.
This does not make sense. You wouldn't send such a prompt to the Claude model. And when you're sending the prompt (anywhere) you don't have the response yet. This is not how distillation works.
What you say makes sense, but further adds to my confusion as to why those model names would appear in input sent to Claude at all, then. EDIT: I guess it might be because someone might point Claude at a compatible API, with its model in the URL, which is of interest to them.
It's unclear to me if they do this to arbitrage on API costs or to steal data, but it makes sense that Anthropic would be interested in detecting when traffic is being proxied to them.