Microsoft introduced two new in-house AI models this week, presenting them as affordable options compared to OpenAI’s offerings. The company also released internal data demonstrating that its models now drive key products across its ecosystem.
Two models, opposite ends of the cost spectrum
The first, MAI-Image-2.5-Pro, is Microsoft’s most advanced image generator to date. It focuses on high-end use cases such as detailed editing and accurate in-image text rendering, an area where many generative AI tools struggle. Pricing is set at $5 per million text input tokens, $8 per million image input tokens, and $106 per million image output tokens.
According to Microsoft, the base version of this model recently ranked second for image editing on Arena, a community leaderboard. Rob Reilly, global chief creative officer at advertising firm WPP, described it as “a strong leap forward for GenMedia tools” in the announcement.
The second model, MAI-Voice-2-Flash, is designed for large-scale enterprise applications like call centers and real-time speech processing. It operates twice as fast as its predecessor and costs 32% less, priced at $15 per million characters. The company optimized it for situations where speed and affordability outweigh subtle expressiveness.
Internal use reveals significant cost reductions
Deployment data shared by Microsoft highlights the impact. In PowerPoint, MAI-Image-2.5 reduces GPU expenses by up to 84% compared to OpenAI’s GPT-Image-2. In OneDrive, the model serves as the default for core image-editing tasks, where it has increased save rates by 26%, lowered latency by 25%, and improved efficiency 2.5 times under medium workloads.
Related: EV charging platform gets 1.5m boost
For voice applications, MAI-Voice-2-Flash now supports Trends 365 Contact Center, which serves companies like T-Mobile and EasyJet. Microsoft states the model cuts GPU costs by up to 89%. It is also part of Azure Voice Live, enabling developers to build speech-to-speech agents.
A key implementation is in healthcare. Microsoft’s Dragon Copilot,
A strategy based on continuous improvement
Microsoft outlined its method in a separate post, describing a feedback system where models evolve through real-world use.
Satya Nadella’s perspective on scalable AI
Microsoft CEO Satya Nadella discussed the announcements in a post on X titled “Frontier Diffusion & Control.” He explained that the company can now deliver advanced capabilities at scale and lower cost through models tailored for high-usage products. “We’re beginning to route traffic across our first-party surfaces to MAI whenever our models match or outperform alternatives,” he wrote.
Nadella acknowledged that OpenAI and Anthropic’s models remain part of Microsoft’s system. However, the implication is clear: Microsoft aims to control the ecosystem, treating third-party models as interchangeable parts. This shift follows reports that Microsoft’s exclusive license to OpenAI’s technology became non-exclusive last year.
Related: Westfalia half-pop bus absorbs luxury RV into van
Mixed reactions from developers and critics
Online responses varied. Some developers welcomed the focus on task-specific models. One user on X asked, “Why use an all-purpose model just to edit a field in Excel?” Others questioned Microsoft’s approach. A designer argued, “Microsoft rarely listens to user feedback,” suggesting the company risks falling behind in the AI race by repeating past missteps.
The criticism has merit. Microsoft’s metrics come from its own evaluations, not independent tests. Yet the company’s argument doesn’t depend on a single figure. Nadella framed AI as a business with tangible costs. When features run on every keystroke for a billion users, an 84% reduction in GPU expenses can determine whether a product is sustainable.
Selling the approach, not just the technology
Microsoft isn’t only using these models internally. It is offering the entire method as an Azure product. Through Foundry and Frontier Tuning, enterprises can train specialized models using their own data and reinforcement learning environments. The message is clear: run AI workloads on Microsoft’s cloud, regardless of where the models originate.
The company is also emphasizing data origins. In an industry facing scrutiny over training sources, Microsoft bets that enterprises and courts will value transparency. “This isn’t the end,” the company stated. “We’re just beginning.”
As AI adoption grows, the ability to customize models for specific needs—like those in fraud detection studies—could become a deciding factor for businesses.
