Microsoft AI is prioritizing token efficiency and small specialist models over general-purpose frontier models, as outlined by AI CEO Mustafa Suleyman. The company trains compact models for single fields, with its latest cybersecurity model, MAI-Cyber-1-Flash, topping the CyberGym benchmark by 12 percentage points over Anthropic's Mythos at half the cost. This result requires the MDASH system, which orchestrates several models and still routes hard tasks to OpenAI's reasoning models. Microsoft also states that MAI-Image-2.5-Flash cuts GPU costs by up to 84 percent compared with GPT-Image-2. Suleyman aims for swappable models to prevent reliance on one model family, though whether the small MAI models partly replacing OpenAI can match its performance remains doubtful. Competition is shifting from individual models to harnesses—software that routes tasks and supplies context—with orchestrators sending most work to cheaper specialists and reserving frontier models for hard cases. Anthropic modeled this approach for Claude Fable 5, while Sakana built Fugu around it.