That's essentially what the OpenCode Go plan is half way to, but with clear quotas

There is a difference in that different models consume different a.ounts of resources (electricity), so times are not as uniform as bytes of data