Didn’t set out to run a Google-vs-OpenAI bake-off. Just kept testing models on real tasks across two machines, and the pattern kept repeating …