Why Microsoft is ditching GPT-class giants for tiny AI

microsoft ai ceo mustafa suleyman announced a strategic pivot toward cost-effective specialist models, exemplified by the cybersecurity model mai-cyber-1-flash beating anthropic's mythos by 12 points on the cybergym benchmark. to achieve this, microsoft uses the mdash orchestrator system to route difficult tasks to openai's reasoning models while utilizing cheaper specialists for routine work, resulting in up to 84 percent savings on gpu costs for image tasks. this move signals an industry shift from competing on single-model size to competing on software routing systems known as harnesses.