which hosted model is actually worth paying for

[mangoquill]

That matches what I was seeing. I ended up in about the same place using vLLM. I got better results once I stopped chasing one metric and just looked at whether the workflow felt usable.