3.1-pro is still very capable, and API is at competitive price vs e.g. Anthropic, they just can't seem to figure out RLHF and harness. It needs a lot of guiding, it tends to be lazy and poorly sticking to instructions by default.

It just feels like many google products really, they are capable of really amazing things, it's just that nobody there seem to care. I would guess they are likely optimizing more for internal use than their vast userbase.

They optimize for making their SRE's lives easier, over quantizing models regardless of how negative an effect that has on the user.