The usage data looks like a classic case of the drift principle. When a model gets heavily optimized for alignment, polish, and safety, you gain consistency but lose some fidelity to the actual task. Newer models think longer, act less, and smooth over edges that used to be useful for real work. Older models aren’t smarter, they’re just sitting earlier on the drift curve, before over compression starts eroding decisiveness. So the specialization we’re seeing may just be developers picking the version where fidelity holds up best, not the one with the highest benchmark score.