March 2026 AI: GPT-5.4 got 1M context + 83% on Humanity's Last Exam. Meanwhile Qwen3.5-9B runs on phones and beats 120B models. Scale isn't the only answer anymore - efficiency is the new moat.