Tech Beetle briefing US AI

Microsoft launches new in-house AI models it says cut costs up to 89% versus OpenAI

Essential brief

Microsoft introduced two new in-house AI models, MAI-Image-2.5-Pro and MAI-Voice-2-Flash, now in public preview, aiming to reduce reliance on OpenAI's models. These models are deployed across Micro

Key topics

microsoft in-house ai models costs versus openai AI MAI-Image-2.5-Pro MAI-Voice-2-Flash OpenAI

Key facts

Microsoft launched two new AI models: MAI-Image-2.5-Pro for high-fidelity image generation and MAI-Voice-2-Flash for high-volume speech workloads.

These models are integrated into major Microsoft products, including Bing, PowerPoint, OneDrive, Dynamics 365, Excel, and GitHub Copilot.

Microsoft reports GPU cost reductions up to 89% compared to OpenAI models, improving efficiency and lowering operational expenses.

The company is offering its hill-climbing AI optimization approach as an Azure product for enterprises to train specialized models on proprietary data.

Highlights

MAI-Image-2.5-Pro is priced at $5 per million text input tokens, $8 per million image input tokens, and $106 per million image output tokens.

MAI-Voice-2-Flash operates twice as fast as its predecessor and costs 32% less at $15 per million characters.

Microsoft’s Dragon Copilot uses MAI-Transcribe-1.5, improving transcription accuracy by 50% across 58 languages.

MAI-Code-1-Flash in GitHub Copilot achieves a 10% higher code acceptance rate than comparable models while using fewer tokens.

Microsoft’s hill-climbing methodology enables efficient model training and deployment on older GPUs, reducing hardware dependency.

Why it matters

Microsoft’s introduction of cost-efficient, high-performance in-house AI models signals a shift in the AI industry toward specialized, scalable solutions tailored for enterprise needs. By reducing reliance on third-party frontier models and optimizing for operational costs, Microsoft aims to make AI integration more sustainable across its vast product ecosystem. This strategy could influence how AI services are developed, deployed, and monetized in the future, emphasizing model independence and cost control.

Microsoft AI released two new in-house models into public preview on July 23, 2026: MAI-Image-2.5-Pro, its highest-fidelity image generator to date, and MAI-Voice-2-Flash, a speech model designed for high-volume enterprise workloads. These launches come about a year after Microsoft committed to developing purpose-built AI models internally. The models are now integrated into key Microsoft products including Bing, PowerPoint, OneDrive, Dynamics 365, Excel, GitHub Copilot, and Azure, marking a shift from research projects to production infrastructure serving millions of users.

MAI-Image-2.5-Pro targets premium image generation tasks such as hero imagery, detailed editing, and precise in-image text rendering, addressing a common challenge in image generation. Pricing for this model is set at $5 per million text input tokens, $8 per million image input tokens, and $106 per million image output tokens. The base MAI-Image-2.5 model recently ranked second on Arena’s image editing leaderboard, indicating strong industry interest.

Conversely, MAI-Voice-2-Flash focuses on cost-effective, high-speed speech processing for applications like call centers and real-time voice agents. It operates twice as fast as its predecessor, MAI-Voice-2, and costs 32% less at $15 per million characters. This model powers Dynamics 365 Contact Center, used by clients such as T-Mobile and EasyJet, and is integrated into Azure Voice Live for developer use.

Microsoft reports substantial cost reductions with these models: MAI-Image-2.5 reduces GPU costs by up to 84% in PowerPoint compared to OpenAI’s GPT-Image-2, and MAI-Voice-2-Flash cuts GPU costs by up to 89% in Dynamics 365 Contact Center. In OneDrive, MAI-Image-2.5 improved save rates by 26%, lowered latency by about 25%, and increased efficiency 2.5 times under medium workloads. Additionally, Microsoft’s Dragon Copilot, used by 170,000 medical providers, now runs on MAI-Transcribe-1.5, improving transcription accuracy by 50% across 58 languages.

Microsoft also detailed its "hill-climbing machine" methodology, which integrates data, models, and product environments to optimize AI performance. For example, MAI-Code-1-Flash, a lightweight coding model in GitHub Copilot, achieves a 10% higher code acceptance rate than comparable models while using fewer tokens. Further training in an Excel environment produced a model comparable to GPT-5.6 for common spreadsheet tasks but capable of running on older GPUs, improving deployment economics.

CEO Satya Nadella framed these developments as part of a strategic approach called "Frontier Diffusion & Control," emphasizing the use of optimized in-house models for routine tasks while reserving frontier models for cutting-edge needs. This approach aims to reduce costs and increase control over AI infrastructure. Microsoft is also offering its hill-climbing approach as a product through Azure, enabling enterprises to train specialized models on proprietary data.

The announcement reflects Microsoft’s evolving AI strategy, balancing partnerships with OpenAI and Anthropic while expanding its own model portfolio. While some developers welcome the cost and performance benefits of task-specific models, others remain cautious about Microsoft’s responsiveness to user feedback and the transparency of its internal metrics. Nonetheless, the move underscores the importance of cost-efficient AI deployment at scale across enterprise applications.

Key topics in this update include microsoft, in-house ai models, and costs.