model routing appears to be very good at protecting 2.5-pro quota. What's the catch?

[timrichardson]

A few hours ago I enabled model routing in my stable release. It seems to be an exceptionally good feature. In my current session, it shows 69 requests to 2.5-pro, 21 to 2.5-flash and 563 to 2.5-flash-lite. It is definitely protecting my quota of pro queries, and I don't notice any decline in usefulness, with a fairly large python code base.

/stats model reports 347 errors for the 563 flash-lite requests, but I see no evidence of that in my experience.

After a few hours, this looks like a transformative feature, consider that by now I would have been well out of my AI Pro quota which is I think 200 2.5 queries.

I am surprised this feature is not on by default. Or I have failed to detect some downside of it.