One could argue that subscription based inference might differ from per-token billed API usage.