Microsoft Rolls Out Two New In-House AI Models, Says They Cost Up to 89% Less Than OpenAI's

Microsoft AI put two new models into public preview on Wednesday, and the bigger story isn't the models themselves. It's what Microsoft says they cost compared to relying on OpenAI.

The releases, MAI-Image-2.5-Pro and MAI-Voice-2-Flash, come from Microsoft's Superintelligence team, according to VentureBeat. That's the group Microsoft built roughly a year ago specifically to stop depending entirely on OpenAI's frontier models for its own products.

This isn't a lab demo. Microsoft says these models already run inside Bing, PowerPoint, OneDrive, Dynamics 365, Excel, GitHub Copilot, and Azure. That's real production traffic, not a press release promise.

Two Models, Two Jobs

MAI-Image-2.5-Pro is the premium option: high-fidelity images, detailed editing, and accurate text rendering inside generated images, a problem that has dogged image generators across the industry for years. Microsoft priced it at $5 per million text input tokens, $8 per million image input tokens, and $106 per million image output tokens.

The base version, MAI-Image-2.5, already ranked No. 2 for image editing on Arena, the community leaderboard widely used as an informal scoreboard for generative media models, according to VentureBeat.

Rob Reilly, global chief creative officer at WPP, one of the world's largest advertising firms, called the Pro model "a strong leap forward for GenMedia tools" and said "Microsoft has firmly established itself among the leaders in generative AI," in a statement included in Microsoft's own announcement.

MAI-Voice-2-Flash targets the opposite end of the market. It runs twice as fast as its predecessor, MAI-Voice-2, and costs 32% less, at $15 per million characters, according to Microsoft. It was first previewed at Microsoft's Build conference earlier this year. The target use case is blunt: call centers, voice agents, and real-time speech systems where cost-per-call and latency matter more than how expressive the voice sounds.

The Cost Argument

Microsoft's pitch, in its own words from the announcement blog: "Each of these enhancements is a step toward the same goal: Microsoft products, powered by Microsoft models."

That's a direct shot at OpenAI, the company Microsoft has invested billions of dollars in and has depended on for years to power Copilot and other AI features. Microsoft is now telling enterprise customers, and implicitly telling OpenAI, that it doesn't need to keep paying OpenAI's prices when its own models can do comparable work for a fraction of the cost. Microsoft claims savings of up to 89%.

VentureBeat's report doesn't include an independent, third-party breakdown validating the 89% figure against specific OpenAI model pricing and equivalent workloads. Enterprise buyers evaluating this claim will want to see apples-to-apples cost comparisons on their own workloads before assuming the savings hold up everywhere.

The Fair Counter-Argument

Skeptics of Microsoft's in-house push have a reasonable point: cheaper isn't automatically better, and a model that costs less per call or per image doesn't matter much if quality drops, especially for high-stakes enterprise use like customer-facing chatbots or brand imagery. Arena's leaderboard ranking is a useful signal, but it's a community benchmark, not a comprehensive industry standard, and a No. 2 ranking in one category (image editing) doesn't confirm Microsoft's models match OpenAI's frontier models across the board on reasoning, coding, or general-purpose tasks.

Microsoft hasn't published head-to-head benchmarks against specific OpenAI models like GPT-4 or GPT-5 series in this announcement, at least not according to what's in VentureBeat's report. That's a gap enterprise IT buyers will likely press Microsoft on before shifting workloads away from OpenAI's models entirely.

What Happens Next

Microsoft still has a deep financial relationship with OpenAI, including a multibillion-dollar investment and cloud-computing commitments running through Azure. Nothing in this announcement signals Microsoft is cutting OpenAI out of Copilot or its core products immediately.

Microsoft has spent a year building a family of purpose-built models, tuned for specific jobs like high-volume voice or premium image generation, rather than betting everything on one frontier model from an outside partner. The next test will be whether Microsoft publishes independently verifiable benchmarks against OpenAI's actual current models, and whether enterprise customers currently paying for OpenAI-powered Copilot features actually migrate to Microsoft's cheaper in-house alternatives once pricing and quality are compared side by side.