Microsoft Rolls Out MAI-Image-2.5-Pro, Accelerates Retreat From OpenAI and Anthropic

Microsoft has released two new models from its in-house MAI family into public preview: MAI-Image-2.5-Pro for image generation and editing, and MAI-Voice-2-Flash for speech synthesis. Alongside the launch, the company published hard numbers on the savings it's getting by replacing OpenAI and Anthropic models with its own technology across its products.

What Was Unveiled on July 23

According to Microsoft, MAI-Image-2.5-Pro is the company's most precise image model to date, standing out for accurate text rendering inside generated graphics and for editing driven by natural-language commands. The model has rolled out in full to Bing Image Creator, as well as to PowerPoint and OneDrive, where it's used to create and touch up images. MAI-Voice-2-Flash is a trimmed-down version of the speech model, optimized for speed and scale in high-volume enterprise use cases. Both models are available in Azure AI Foundry and in the MAI Playground.

Deployment Numbers

Rather than offer general assurances of progress, Microsoft this time put out specific figures. In Dynamics 365 Contact Center, where voice models handle customer calls, switching to in-house technology cut GPU costs by 89 percent. In PowerPoint, the reduction reached 84 percent, and in OneDrive, moving to MAI-Image raised the rate at which edited images get saved by 26 percent. The Dragon Copilot transcription model, used for medical documentation, now supports 58 languages.

Turning Away From OpenAI and Anthropic

These launches are the next step in a strategy Microsoft kicked off in June 2026 at its Build conference, where it unveiled seven models in the MAI family: the flagship reasoning model MAI-Thinking-1, the coding model MAI-Code-1-Flash with 5 billion active parameters, two image models, a transcription model, and two voice models. Since July 7, 2026, Microsoft has actually begun shifting a portion of Copilot traffic in Excel and Outlook to these models, which now handle tens of thousands of queries a week, mostly routine tasks like summarizing emails or formatting spreadsheets.

Microsoft AI CEO Mustafa Suleyman spelled out the motivation directly, pointing to the cost of working with Anthropic, whose models Microsoft has so far used in parts of its office applications.

Anthropic is extremely expensive, and I think a lot of people are urgently looking for alternatives. We're paying Anthropic a lot of money, so our goal is to reduce, and ultimately eliminate, that cost - Mustafa Suleyman, CEO, Microsoft AI

This isn't a full break yet. OpenAI's models still handle more complex, demanding tasks across Microsoft's products, and Anthropic's models remain built into select features of select office apps. Microsoft does stress, however, that it develops its image and voice models without distilling from third-party technology, meant to signal the company's growing technological independence, even though Microsoft has invested roughly $13 billion in OpenAI since 2019 and holds exclusive cloud rights to its models.

What It Means for Customers

For enterprise customers using Microsoft 365 Copilot, this means some of the responses generated in Excel, Outlook, or PowerPoint are already coming from MAI models rather than OpenAI or Anthropic, though Microsoft hasn't disclosed in detail which specific query types are being redirected. For companies billed for AI usage by token cost, the lower rates on MAI-Image-2.5-Pro and MAI-Voice-2-Flash could meaningfully cut Copilot feature bills, if Microsoft passes those savings on through licensing prices. Cutting orders for OpenAI and Anthropic models across products with billions of users also hits the inference revenue of both providers, for whom Microsoft is one of the largest customers.

MAI-Image-2.5-Pro is a big step forward for GenMedia tools - Rob Reilly, Global Chief Creative Officer, WPP