Microsoft’s internal AI development is starting to show results, with the MAI family of models and the Maia inference platform — both designed to lower token costs.