Yep, but even 0.006 is quite a big improvement. I'm curious now to test the model on some token-heavy tasks, like code exploration before a coding session, to see whether it will decrease the total cost of the task in the end or not.