Microsoft's New In-House AI Models: MAI-Image-2.5-Pro and MAI-Voice-2-Flash Cut GPU Costs Up to 89%

The two models launched

• MAI-Image-2.5-Pro: Microsoft's highest-fidelity image generation model to date, built for professional visuals, detailed editing, and precise text rendering inside generated images . It is priced via Microsoft Foundry at $5 per 1M text input tokens, $8 per 1M image input tokens, and $106 per 1M image output tokens .
• MAI-Voice-2-Flash: A low-latency text-to-speech model optimized for speed and scale in customer-service and real-time voice applications . Microsoft says it is twice as fast and 32% more cost-effective than the standard MAI-Voice-2 version .

Both models entered public preview through Microsoft Foundry on July 24 .

Where they are deployed

Microsoft is rolling the new models into several flagship products immediately :

• Bing Image Creator: Now 100% in-house, running end-to-end on MAI-Image-2.5-Pro . The previous version still depended on OpenAI's image models.
• PowerPoint: MAI-Image-2.5-Pro has replaced OpenAI's GPT-Image-2 as the default image generation engine .
• OneDrive: The image model is now the default tool for AI-generated images .
• Dynamics 365 Contact Center: MAI-Voice-2-Flash powers real-time voice and customer-service workflows . Microsoft AI CEO Mustafa Suleyman noted integrations at customers including T-Mobile and EasyJet .
• Microsoft Foundry (Azure AI Foundry): Both models are available for developers in public preview .

Microsoft states that MAI models are now deployed across more than half of its products, with further replacements underway in Excel and Outlook .

GPU cost reductions vs. OpenAI models

Microsoft provided specific GPU compute cost reduction percentages — meaning the savings on inference infrastructure when running its own models versus licensing or hosting OpenAI's equivalents. These are not API pricing differences but actual hardware-level savings :

These reductions are significant enough to change Microsoft's unit economics for AI features across its Office and cloud suite.

How the models fit into Microsoft's MAI family

The July 24 releases are specialized variants within a broader seven-model MAI family announced at Microsoft Build in June 2026 :

1. Reasoning: MAI-Thinking-1
2. Image: MAI-Image-2.5, MAI-Image-2.5-Flash, and now MAI-Image-2.5-Pro
3. Voice / Speech: MAI-Voice-1, MAI-Voice-2, and MAI-Voice-2-Flash
4. Transcription: MAI-Transcribe-1 / MAI-Transcribe-1.5
5. Coding: MAI-Code-1-Flash

MAI-Image-2.5-Pro adds a premium tier for maximum image quality, while MAI-Voice-2-Flash adds a speed-optimized tier for high-volume, low-latency voice . The entire family is trained on Microsoft's own Maia 200 custom AI accelerators without distillation from third-party models, with a claimed 1.4x efficiency advantage from the hardware-software co-design .

Strategy: Reducing reliance on OpenAI and third-party models

Microsoft is executing a deliberate multi-phase strategy to build full-stack AI self-sufficiency :

• Replace third-party models first-party: PowerPoint, Bing Image Creator, and OneDrive have all switched from OpenAI's image models to MAI-Image-2.5-Pro . This is described as a "minor shift away from OpenAI" that is accelerating .
• Keep OpenAI as a partner, not a sole supplier: Microsoft remains an investor in OpenAI and Azure hosts OpenAI models, but the company is no longer dependent on OpenAI for its own first-party products. The goal is a multi-model strategy with competitive alternatives available .
• Cost control as strategic leverage: The 84–89% GPU cost reductions mean Microsoft can offer competitive AI features at far lower per-user cost, insulating its margins and strengthening its negotiating position with all external model providers .
• Build full-stack vertically: Own chips (Maia 200), own models (MAI family), own inference infrastructure (Azure), and own product surface area (Office, Bing, Dynamics, Windows) .

Cost, pricing, and self-sufficiency implications

• Infrastructure cost elimination: Running first-party models on Maia 200 chips eliminates the margin paid to third-party model providers for every inference .
• Durable AI features: The dramatic GPU cost reductions suggest the MAI models are purpose-built for Microsoft's specific product use cases rather than being general-purpose foundation models, making AI features in Office and Dynamics more resilient to OpenAI price changes .
• Bottom line: The July 24 launch completes the replacement of OpenAI image and voice models in Microsoft's highest-volume consumer and enterprise products at a fraction of the GPU cost . Microsoft is building an independent, vertically integrated AI stack that leaves the company far less exposed to any single external dependency.