It's only benchmarks where I hear about Google's models being frontier. Never really get to use it because they don't have a usable subscription model on par with Claude/Openai ones in terms of usage.