Microsoft Moves Bing and PowerPoint Image Features to MAI Models

Microsoft said on July 23 that Bing Image Creator is now fully powered by its MAI-Image-2.5 model, while PowerPoint and OneDrive also use the model for image editing, reporting up to 84% lower GPU costs than GPT-Image-2 in PowerPoint. The shift from OpenAI models to in-house MAI models is driven by cost and product fit, with Microsoft AI chief Mustafa Suleyman telling Bloomberg the internal technology is faster, cheaper, and higher quality.

Microsoft Moves Bing and PowerPoint Image Features to MAI Models Microsoft said on July 23 that Bing Image Creator is now powered end to end by MAI-Image-2.5, while PowerPoint uses the model for image-to-image features and OneDrive uses it for key editing workflows. The company reported up to 84% lower PowerPoint GPU costs than GPT-Image-2, highlighting cost and product fit as drivers of its shift from OpenAI image models. Microsoft said on July 23 that its own MAI image models are now running across Bing Image Creator, PowerPoint and OneDrive, extending the company's move from third-party models to systems built for specific product workloads. Bloomberg separately reported that Microsoft is replacing OpenAI image-generation technology in PowerPoint and Bing. Where MAI-Image-2.5 is running Microsoft's announcement says Bing Image Creator is now fully powered by MAI-Image-2.5 for generation and editing. In PowerPoint, the model is in production for image-to-image features, which modify or extend an existing image from a prompt. Microsoft reported that those PowerPoint workloads use up to 84% less GPU capacity than GPT-Image-2. OneDrive also uses MAI-Image-2.5 as the default model for key image-editing scenarios. Microsoft said the rollout increased the share of edited images that users save by 26%, reduced P95 latency by about 25%, and delivered 2.5 times greater efficiency under medium-utilization workloads. The same announcement introduced MAI-Image-2.5-Pro in public preview for higher-fidelity generation and editing. That variant expands the product family; it is distinct from the production MAI-Image-2.5 deployments behind the reported Bing, PowerPoint and OneDrive results. Cost becomes a product decision Microsoft AI chief Mustafa Suleyman told Bloomberg that the in-house image technology is faster, cheaper and higher quality, and that it improves retention. Bloomberg reported that Microsoft can use OpenAI models without a model-access fee under the companies' partnership, but Microsoft still pays for the computing infrastructure required to run them. A more efficient internal model can therefore reduce serving costs even when model access itself is not billed. The performance figures are Microsoft claims, not independent benchmarks. They are still useful because they describe production measures that model leaderboards often omit: GPU cost per workload, tail latency and whether users keep the output. What the shift does and does not show The deployment does not establish that MAI-Image-2.5 is better than every OpenAI model for every image task, nor does it mean Microsoft's broader OpenAI partnership has ended. It shows a narrower operational choice: Microsoft is assigning high-volume image features to an internal model that it says better fits those products' quality, latency and cost requirements. For teams deploying generative media, the practical lesson is to evaluate models at the level of the actual workflow. A model that meets the required quality bar with lower accelerator cost and latency can be the stronger production choice even if another model leads on a general benchmark. Key Points - 1Microsoft said Bing Image Creator is now powered end to end by MAI-Image-2.5 for image generation and editing. - 2PowerPoint uses MAI-Image-2.5 for image-to-image features, with Microsoft reporting up to 84% lower GPU costs than GPT-Image-2. - 3OneDrive uses the model for key editing workflows, where Microsoft reported higher save rates, lower P95 latency and greater serving efficiency. Scoring Rationale The deployment affects widely used Microsoft products and supplies unusually concrete production claims about GPU cost, latency, efficiency and user behavior. Those metrics make the event relevant to teams choosing models for high-volume generative-image features. Sources Primary source and supporting public references used for this report. Practice interview problems based on real data 1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with. Try 250 free problems /problems