I think they're concentrating on users and not benchmarks. Gemini is still the best model at plain old searches, like, for parsing images or "what's that poem with this and this in it". It seems to have much more breadth of reasoning than any of the other models, e.g. Gemini 3.1 is still the top model on simple-bench other than Fable.

New 3.5 Flash model is clearly quite optimized for speed. People on this site just look at benchmarks, and underestimate how vital speed is to user experience. No other free model comes close in terms of speed and intelligence, so it seems like the correct strategy to slowly win the mass market, as almost everyone will form their experience of AI through its free version.