> which OpenAI seems to allow via API but not Anthropic

Does openai still allow logprobs in their current gen models?

It's not that light on the details. I read the paper and they train several different models in parallel over a few different domains and then they distill from their own models to get the final model.