Microsoft expands in-house AI push with cheaper models for image and voice tasks

Microsoft is expanding its in-house AI portfolio with new image and speech models, while shifting more product traffic away from expensive general-purpose systems.

The company has placed MAI-Image-2.5-Pro and MAI-Voice-2-Flash into public preview. MAI-Image-2.5-Pro targets detailed image creation, editing and text within images. MAI-Voice-2-Flash is designed for large-scale speech applications such as contact centres and voice agents.

Microsoft says the releases support a broader strategy: use smaller, specialised models for frequent product tasks, while reserving frontier models from partners such as OpenAI and Anthropic for more demanding work. CEO Satya Nadella describes this as improving the “cost-to-outcome” ratio for real-world software.

Lower costs for high-volume AI features

Microsoft reports that its internally developed models are already operating across products including Bing, PowerPoint, OneDrive, Dynamics 365, Excel and GitHub Copilot. The company says MAI-Image-2.5 cuts GPU costs in PowerPoint by up to 84% compared with OpenAI’s GPT-Image-2. In Dynamics 365 Contact Center, it says MAI-Voice-2-Flash reduces GPU costs by up to 89%.

These figures come from Microsoft’s own production measurements and have not been independently verified.

The company also says its MAI-Code-1-Flash model achieves a roughly 10% higher code acceptance rate in VS Code than GPT-5.4 Mini and Claude Haiku 4.5, while using fewer tokens. Microsoft further trained this model for Excel workflows and says user feedback puts it on par with GPT-5.6 for common spreadsheet tasks.

Microsoft argues that efficient models can run on Nvidia H100 and A100 hardware, reducing serving costs and preserving newer chips for training. It plans to offer its training and evaluation approach through Microsoft Foundry, allowing organisations to develop specialised models using their own workflows and data.

Sources

• Microsoft launches new in-house AI models it says cut costs up to 89% versus OpenAI – VentureBeat
• Hill-Climbing MAI models for GitHub Copilot and Excel – Microsoft AI
• In a world where software has real marginal cost for the first time, how do we ensure frontier benefits are diffused across the entire ecosystem? – LinkedIn