Microsoft's own AI models are taking over more of the products that small businesses already pay for. On 23 July 2026, Microsoft AI, the company's in-house model lab, put two new models into public preview: MAI-Image-2.5-Pro for high-quality images and MAI-Voice-2-Flash for fast, lower-cost speech. In the same post it confirmed that its MAI models now run image tools in Bing, PowerPoint and OneDrive and the voices in Dynamics 365 Contact Center. If your business uses Microsoft software, some of the AI features your staff and customers meet are now built on Microsoft's own technology.
What happened
The news came from Microsoft AI's Superintelligence team in a post introducing MAI-Image-2.5-Pro and MAI-Voice-2-Flash. MAI is the name Microsoft gives the models it builds itself. The post has two parts.
The first is two new model variants, both in public preview for developers through Microsoft Foundry and available to try in the MAI Playground:
- MAI-Image-2.5-Pro is aimed at work where quality matters most. Microsoft lists hero imagery, detailed editing and precise text inside an image, and calls it "our highest-fidelity image model to date".
- MAI-Voice-2-Flash turns text into speech and is built for speed and volume. Microsoft says it is twice as fast as MAI-Voice-2 and 32 per cent cheaper, while keeping the natural rhythm and sound quality of MAI-Voice-2.
The second part matters more to most businesses: a list of Microsoft products that have switched to MAI models. Bing Image Creator now uses MAI-Image-2.5 by default from start to finish. PowerPoint uses it for editing one image into another. OneDrive uses it for its main image-editing features. Dynamics 365 Contact Center, Microsoft's platform for building call centre agents, now speaks with MAI-Voice-2-Flash. Microsoft also says its transcription model, MAI-Transcribe-1.5, now supports the multilingual workflow in Dragon Copilot, a product used by medical providers, across 58 languages.
Microsoft describes these as models "trained on clean, traceable, enterprise-grade data, without distillation from third-party models". Distillation means training one model to imitate the answers of another, so the claim is that these models were not copied from a competitor's.
Key details
| Item | What Microsoft announced |
|---|---|
| MAI-Image-2.5-Pro | Public preview. Priced at US$5 per million text input tokens, US$8 per million image input tokens and US$106 per million image output tokens. |
| MAI-Voice-2-Flash | Public preview. Priced at US$15 per million characters. Twice as fast as MAI-Voice-2 and 32 per cent cheaper. |
| Bing Image Creator | MAI-Image-2.5 is the default model for generating and editing images. |
| PowerPoint | MAI-Image-2.5 is in production for image-to-image features. Microsoft reports GPU costs up to 84 per cent lower than with GPT-Image-2. |
| OneDrive | MAI-Image-2.5 is the default for key image-editing scenarios. Microsoft reports that save rates rose 26 per cent and that response times at the slow end (the 95th percentile) are about 25 per cent shorter. |
| Dynamics 365 Contact Center | Voices are now produced by MAI-Voice-2-Flash, with GPU costs up to 89 per cent lower. |
| Azure Voice Live | MAI-Voice-2-Flash is available for developers building speech-to-speech voice agents. |
A token is the unit these services bill by: a small piece of text or image data. A GPU is the specialised processor that AI models run on, and it is the main running cost of any AI feature. Every figure in the table is Microsoft's own, and the post does not include independent benchmark results for the two new models.
Why it matters
Microsoft is becoming its own supplier. The post sums up the direction as "Microsoft products, powered by Microsoft models". The PowerPoint comparison is telling, because the model it measures against, GPT-Image-2, is an OpenAI model. Microsoft says it evaluates every available model before choosing one for a product, and that its own models are now winning more of those comparisons on quality, speed and efficiency. Windows Report's coverage of the launch notes that MAI-Image-2.5 was only unveiled in May, so the move from announcement to default has been quick.
The savings are Microsoft's, at least for now. The 84 and 89 per cent figures describe the GPU cost of running these features. The announcement says nothing about the price of Microsoft 365, Dynamics 365 or any other subscription. Lower running costs make it easier for a vendor to include AI features in existing plans or to offer more generous limits, but that is a possibility, not something Microsoft has promised.
Models now come in families. Microsoft's reasoning is that a creative studio and a busy call centre need different trade-offs, so it offers a Pro variant where quality comes first and a Flash variant where speed and cost come first. The same pattern is appearing across the industry, and it is useful for buyers: the question is no longer "which AI" but "which grade of AI is enough for this job".
Synthetic voices are entering ordinary customer service. Microsoft names T-Mobile and EasyJet as customers of the contact centre platform that now uses its voice model. Natural-sounding automated voices are no longer an experiment. Customers will increasingly meet them on the phone, including when they ring smaller businesses that use the same software.
What this means for businesses
For most small and medium businesses there is nothing to buy or install. The changes to Bing, PowerPoint and OneDrive arrive inside the products. What changes is the output: images produced or edited in those tools may look different from what staff were getting before, and editing instructions may behave differently.
Two habits are worth setting now. First, treat generated images as drafts. Microsoft says the Pro model renders text inside images precisely, but a price, a phone number or a product name still needs a person to check it, and a generated picture of a product, a finished job or a person that does not exist can mislead a customer. Second, be open about automated voices. The Office of the Australian Information Commissioner says in its guidance on commercially available AI products that public-facing AI tools should be clearly identified as such to the people using them, and that privacy obligations apply to personal information put into an AI system. A phone call is full of personal information.
For businesses with their own software, the two preview models are priced per use and available through Foundry. A preview can change before general release, so it suits testing more than anything the business depends on.
A short checklist:
- Ask whoever manages your Microsoft subscriptions which AI image features are switched on for staff.
- Decide who checks AI-generated images before they go to a client or the public.
- If you use, or are considering, an automated voice on your phone line, decide how callers are told and how they reach a person.
- Keep customer personal information out of AI tools unless the business has approved that use.
- If your own systems generate images or speech, ask your developer whether the new price points change the sums.
Comingwave is a technology company that provides managed IT support, business systems and integrations and custom software to small and medium businesses. If you would like help reviewing how these tools are set up in your business, request a free first consultation. We reply within one business day.
Key takeaways
- On 23 July 2026 Microsoft put MAI-Image-2.5-Pro and MAI-Voice-2-Flash into public preview.
- Microsoft's own models now run image features in Bing Image Creator, PowerPoint and OneDrive, and the voices in Dynamics 365 Contact Center.
- The large cost reductions Microsoft reports are its own running costs. No subscription price changes were announced.
- Generated images and automated voices still need a human check and clear disclosure to customers.
Frequently asked questions
What are Microsoft MAI models?
MAI is the name for the AI models that Microsoft AI builds in-house. The models named in the announcement cover images, voice and transcription. Microsoft is using them inside its own products and is offering the two new ones to developers in preview.
What is the difference between MAI-Image-2.5-Pro and MAI-Voice-2-Flash?
They do different jobs. MAI-Image-2.5-Pro creates and edits images and is the higher-quality option in Microsoft's image family. MAI-Voice-2-Flash turns text into speech and is the faster, cheaper option in its voice family.
Do Microsoft 365 customers need to do anything?
No action is required. The switch to MAI models in Bing Image Creator, PowerPoint and OneDrive happens inside those products. It is still sensible to review which AI features staff can use and how the results are checked.
Will this make Microsoft subscriptions cheaper?
Microsoft has not said so. The cost reductions in the announcement refer to the GPU cost of running the features, and the post does not mention any change to subscription prices.
Can a small business use these models directly?
Yes, through a developer. Both models are in public preview on Microsoft Foundry with per-use pricing, and Microsoft offers a playground for trying them. Because they are previews, they are better suited to testing than to systems the business relies on.