tl;dr: Small, fast models like "gpt-5.6-luna" and GLM 5.3 have become capable enough to handle complex tasks (research, personalized news aggregation) at costs around $0.10 instead of $1+, potentially unlocking consumer AI apps that were previously uneconomical due to token costs. The author argues that while frontier models will remain essential for "IQ 180" novel problem-solving, most business work is "token spewer" grunt work—responsiveness, follow-ups, coordination—where cheap/good-enough models could soon dominate, pending advances in harnesses, safety, and permissions.