Editorial Team
6 min read
Microsoft has put MAI-Image-2.5-Pro and MAI-Voice-2-Flash into public preview in Foundry. The launch extends the company’s in-house AI options, but its evidence for lower cost and broader deployment is chiefly tied to MAI-Image-2.5 and company-reported workloads—not an independently comparable test of the new image model.
Microsoft, the operator of Azure and a major OpenAI shareholder, has added two in-house generative-AI models to public preview in Microsoft Foundry: MAI-Image-2.5-Pro for image generation and editing, and MAI-Voice-2-Flash for text-to-speech. The release gives developers more choices inside Microsoft’s platform, but the company’s strongest operating evidence belongs to related models already in production—not to an independent test of the new image preview.
That distinction matters because the announcement is also a strategic claim. Microsoft says in its release that it began building purpose-built MAI models in-house a year ago and is now choosing models for individual product surfaces on a quality-speed-cost curve. The available evidence supports a growing internal model portfolio; it does not show that Microsoft has replaced outside models everywhere, or that Image Pro has already proved itself in the same production settings as MAI-Image-2.5.
Microsoft’s announcement calls MAI-Image-2.5-Pro its highest-fidelity image model to date, aimed at detailed generation, editing and in-image text. The Foundry model card identifies it as a preview, diffusion-based text-to-image and image-to-image model. Its stated strengths are controlled edits—such as changing an object or layout, updating text, cleaning artifacts and preserving visual consistency across iterations.
MAI-Voice-2-Flash is a text-to-speech model optimized for low-latency interactions. Microsoft positions it for voice agents, assistants and contact-center or IVR flows; the technical documentation says it supports 15 languages and 18 locales, expressive controls through SSML, and licensed prebuilt voices. A developer can also prompt a voice from a five- to 60-second reference clip, subject to the safeguards below.
| Model | Microsoft’s stated role | Published price | Important qualification |
|---|---|---|---|
| MAI-Image-2.5-Pro | Higher-fidelity image creation and controlled editing | $5 per million text-input tokens; $8 per million image-input tokens; $106 per million image-output tokens | Preview; the pricing separates text input, image input and image output. |
| MAI-Voice-2-Flash | Fast, high-volume text-to-speech | $15 per million characters | Preview; its price unit cannot be compared directly with the image model’s token-based rates. |
Microsoft says in its announcement that Voice Flash is twice as fast and 32% cheaper than MAI-Voice-2 while retaining that model’s natural prosody and acoustic quality. Those are the company’s comparisons between two of its voice models, rather than an independent benchmark or a comparison with a rival service.
Microsoft reports up to 84% GPU-cost reduction for MAI-Image-2.5 in PowerPoint and up to 89% for MAI-Voice-2-Flash in Dynamics 365 Contact Center; company-reported figures for different workloads. Source: Microsoft AI.
Microsoft says MAI-Image-2.5, rather than Pro, now powers Bing Image Creator end-to-end and is used for PowerPoint image-to-image capabilities and key OneDrive image-editing scenarios. In the PowerPoint comparison, Microsoft says MAI-Image-2.5 cut GPU costs by as much as 84% against GPT-Image-2, an image model from OpenAI, the AI company whose models and products Microsoft licenses through 2032 under their amended agreement.
For OneDrive, Microsoft reports in the release a 26% increase in save rates, an approximately 25% reduction in P95 latency and 2.5 times greater efficiency under medium-utilization production workloads after rollout. The measures have different denominators and baselines: save rate concerns a user action, P95 is a tail-latency measure, and efficiency is not defined further in the release. They establish Microsoft’s reported operating outcome for those scenarios, not a portable quality ranking for Image Pro.
The distinction has been blurred in coverage. One report initially described Image Pro as being used in PowerPoint, Bing and OneDrive, but its later wording and Microsoft’s original announcement identify MAI-Image-2.5 for PowerPoint and OneDrive. A separate account of the launch likewise attributes Bing, PowerPoint and OneDrive deployments to MAI-Image-2.5. The retained record therefore does not establish that Pro is serving those product experiences.
Voice Flash does have a named production use: Microsoft says in the release that it powers Dynamics 365 Contact Center, the company’s enterprise platform for building call-center agents, and has reduced GPU costs by up to 89% there. It is also integrated into Azure Voice Live for speech-to-speech agents. Those are again company statements for particular workloads, not an estimate of savings a customer should expect across regions or traffic patterns.
The MAI-Voice family is available through Azure Speech in Foundry Tools. The documentation ties Flash to Azure accounts, Speech resources and regions that support MAI models; it does not supply a general latency figure or a universal deployment-cost calculation.
More importantly, the documentation describes the family as a public preview without a service-level agreement and says it is not recommended for production workloads. Voice prompting requires limited-access approval and consent safeguards. Microsoft says only authorized, licensed voices can be synthesized in production, so a five- to 60-second clip is not a shortcut around the access and consent process.
That creates a meaningful boundary around the release. Microsoft can point to its own Dynamics deployment, while a third-party buyer considering Flash must still assess preview terms, regional availability, consent operations and its own throughput before treating the claimed speed or GPU savings as a budget forecast.
The new previews do broaden Microsoft’s ability to select its own models for workloads it controls. But the company’s April amended agreement with OpenAI preserved a substantial partnership: Microsoft remains OpenAI’s primary cloud partner, OpenAI products ship first on Azure unless Microsoft cannot and chooses not to support the needed capabilities, and Microsoft’s now non-exclusive license to OpenAI intellectual property runs through 2032. Microsoft also says it remains a major shareholder, while OpenAI can serve products across cloud providers.
So the immediate consequence is more model-selection control inside Microsoft’s product and cloud stack, not evidence of a clean break. The financial and infrastructure stakes still run both ways: the companies say they will continue work on new datacenter capacity and next-generation silicon, even as the amended agreement gives each more flexibility.
The next useful evidence is not another product assertion but comparable results for the previews themselves. For Image Pro, that means disclosed evaluation methods and production performance that separate it from MAI-Image-2.5, including the workload and output-token volume behind its $106-per-million-output-token price. For Voice Flash, it means customer deployments that identify region, traffic conditions, latency measurement and the operational effect of its limited-access voice-cloning controls.
Until then, the evidence supports a more precise reading: Microsoft has expanded a credible in-house MAI lineup and is using related versions in its products, but the commercial and technical case for each new preview remains specific to the workload—and, for the headline image model, largely still to be demonstrated.
Xiaomi says its humanoid robot sorted car parts for a full shift and corrected errors autonomously. The result extends its factory trials, but published evidence does not yet establish deployment costs or reliability across repeated shifts.
Railway’s CEO apologized after an engineer’s AI agents targeted rival Primcloud during testing. Primcloud says attempts to obtain customer secrets failed, leaving questions about the agents’ instructions and Railway’s controls.
Cortex Training gives developers custom training APIs on managed GPUs. Snowflake is extending its existing model-customization tools, with enterprise-data integration and a cost claim limited to LoRA fine-tuning.
Apple’s developer preview extends Siri AI beyond predefined app categories and adds guidance for complex tasks. Testing is limited to development builds and internal testers, with App Store submissions planned for spring 2027.
Jill Hazelbaker will join OpenAI as chief corporate affairs officer after a decade at Uber, bringing experience in technology communications and political campaigns to a role focused on reputation, positioning and public trust.
The watchOS 27.3 developer beta brings Live Rewind and Siri Recap to Apple Watch Series 12 and Ultra 4. The opt-in tools offer brief transcripts and conversation summaries, but require a compatible iPhone and carry Apple warnings about accuracy and sensitive information.
Traders reportedly extract Samsung and SK hynix memory from processor packages shipped into China, selling it mainly to Huawei customers. The account highlights the memory constraint behind Huawei’s AI chip rollout, but leaves trade volumes, recovery yields and transaction legality unresolved.
Creative-industry groups are challenging OpenAI and Microsoft’s arguments against licensing AI training material, as a New York court weighs fair use. Their briefs share an emphasis on licensing but differ over what the judge should decide.
Traversal's Series B backs its expansion from diagnosing software outages toward autonomous production operations. Company-reported adoption, cloud competition and unfinished access controls frame the next tests.
The startup formerly called TesterArmy is replacing its testing agent with an open-source framework, betting that adoption will generate demand for paid infrastructure and enterprise services.
Fal has announced Sequoia-led and Bessemer-led funding rounds for its 2027 expansion, adding post-trained media models to an established inference business. The announcements leave the new rounds’ size and comparative serving economics undisclosed.
SK hynix bought existing Penguin Solutions preferred shares from an SK Telecom investment vehicle. The deal brings director nomination rights and a convertible stake in an existing AI infrastructure partner, while sending the purchase price to the seller.
Google plans to switch off Gemini and Search for school-managed K–12 student accounts unless administrators enable them. Schools will gain control over access, but Google says keeping Search without its AI features will not be an option.
Anthropic is expected to exceed 3.5 GW of computing capacity in 2026 and aims to double it in 2027. Its filing shows $895.4 billion in commitments, substantial financial resources and an immediate cost from scarce capacity.
Independent tests reproduce MIT researchers’ repair of Qwen’s repetitive reasoning. The edited model still beats a stronger unedited configuration, but by much less; broader claims about creating capabilities remain unproven.
Gartner forecasts that open-weight models will execute 90% of enterprise operational agent steps by 2030. Its cost estimates favor routine, high-volume work, but infrastructure costs and the choice of comparison model leave enterprise savings to be proved.
FTAI Aviation is financing a planned 100-turbine production run in 2027, while ProEnergy has a 650 MW equipment order from Crusoe. Converted aircraft engines offer another supply route for data-center power, but delivery targets, affiliated financing and unproven operating economics qualify the opportunity.
Meta is marketing dollar bonds in a plan to borrow at least $20 billion, including an expected euro offering. Capital spending exceeded operating cash flow last quarter, but the debt is part of a wider financing strategy that also includes equity offerings and a proposed non-voting share-sale program.
A nonbinding opinion targets real-time retrieval and recognisable reuse of journalism, while rejecting a monopoly over news facts and recommending that questions about model training be dismissed as hypothetical.
A study of Lovable projects found passport details, customer records and private messages accessible without login, raising questions about database permissions and the limits of automated security checks. Its sample does not establish a platform-wide leak rate.