Корпоративный AI3 мин. чтения

Microsoft Unveils Proprietary AI Models Cutting Costs by 89% vs OpenAI

Microsoft has introduced two new in-house AI models — MAI-Image-2.5-Pro for image generation and MAI-Voice-2-Flash for speech processing. The company claims up to 89% cost reduction compared to OpenAI solutions.

Microsoft Unveils Proprietary AI Models Cutting Costs by 89% vs OpenAI

Microsoft has officially launched two new proprietary AI models — MAI-Image-2.5-Pro for image generation and MAI-Voice-2-Flash for speech processing. Already integrated into products like Bing and Dynamics 365, these models demonstrate up to 89% GPU cost reduction compared to equivalent OpenAI solutions. This marks Microsoft's strategic move to reduce third-party AI dependency.

At a glance: Key facts

  • MAI-Image-2.5-Pro delivers enhanced detail in image generation
  • MAI-Voice-2-Flash cuts voice processing costs by 32% for call centers
  • Dynamics 365 Contact Center shows 89% GPU cost savings
  • OneDrive reports 26% increase in file retention after switching to MAI-Image-2.5-Pro
  • Models currently power Bing, PowerPoint, Excel and GitHub Copilot
  • Microsoft employs 'hill-climbing' methodology for performance optimization
  • 50% reduction in speech recognition errors for Dragon Copilot medical transcriptions
  • MAI-Code-1-Flash shows 10% better code acceptance rate than Claude Haiku 4.5
Visual representation of Microsoft's new AI models
Architecture comparison between Microsoft's new models and previous versions

Microsoft's new AI models explained

MAI-Image-2.5-Pro is Microsoft's flagship image generation model specializing in detailed visualization and accurate text rendering within images. It ranked second on Arena, the generative media evaluation platform. Technical features include:

  • 8K resolution support with preserved detail
  • Enhanced semantic object segmentation
  • 92% text rendering accuracy in Microsoft's tests

MAI-Voice-2-Flash handles mass-scale speech processing for call centers and voice agents. The model operates twice as fast as its predecessor at $15 per million characters. Key advantages:

  • 58 language and accent support
  • <200ms latency for 95% of requests
  • Automatic background noise adaptation

Cost efficiency and performance

Dynamics 365 Contact Center (used by T-Mobile and EasyJet) achieved 89% GPU cost reduction with MAI-Voice-2-Flash. OneDrive saw 26% increased file retention and 25% lower latency after adopting MAI-Image-2.5-Pro.

Optimization techniques

Microsoft's 'hill-climbing' methodology continuously improves models using real user data. MAI-Code-1-Flash for GitHub Copilot shows 10% better performance than OpenAI and Anthropic alternatives. Benchmark results:

Model Acceptance Rate Tokens per task Retention Rate
MAI-Code-1-Flash 72% 142 89%
GPT-5.4 Mini 65% 158 83%
Claude Haiku 4.5 62% 161 78%

Current implementations

  • Bing Image Creator processes 4.2M daily requests with MAI-Image-2.5-Pro
  • Dynamics 365 Contact Center handles 12M monthly sessions with MAI-Voice-2-Flash
  • Dragon Copilot serves 28M quarterly patients with MAI-Transcribe-1.5
  • Excel and PowerPoint automate 87% routine tasks using new models
  • Azure Voice Live offers MAI-Voice-2-Flash as developer service

Comparison with OpenAI

While maintaining OpenAI partnerships for complex tasks, Microsoft's new solutions target specific scenarios for greater cost efficiency.

Metric MAI-Image-2.5-Pro GPT-Image-2 (OpenAI)
Cost (per million tokens) $5 (text), $8 (image) 84% more expensive
Latency (P95) 25% lower Higher
GPU efficiency 2.5x better Lower
Language support 32 core languages 45+
Max resolution 8K 4K

Development roadmap

Microsoft plans expanded deployment across Azure and GitHub Copilot, with industry-specific adaptations for healthcare and finance. Expected updates:

  • MAI-Voice integration in Teams for auto-transcription
  • Arabic and Hindi support in MAI-Transcribe by late 2024
  • Optimization for Nvidia Blackwell GPUs
  • Enterprise adoption partner program

Questions & answers

How much cheaper are Microsoft's new models vs OpenAI?

MAI-Image-2.5-Pro reduces GPU costs by 84% versus GPT-Image-2, while MAI-Voice-2-Flash delivers 89% savings in Dynamics 365 Contact Center. Excel automation with MAI-Code-1-Flash saves $1.2M annually per 10,000 users.

Which Microsoft products already use MAI-Image-2.5-Pro?

The model powers Bing Image Creator (4.2M requests/day), PowerPoint Designer, OneDrive auto-retouching, and Excel data visualization.

How does MAI-Voice-2-Flash improve call centers?

Processing voice requests twice as fast with <200ms latency, critical for high-volume calls. In 1M session tests, recognition accuracy reached 94.7% vs competitors' 91.2%.

Why is Microsoft transitioning to proprietary AI models?

The company aims to reduce costs (projected $220M annual savings) and gain technology control. Proprietary models avoid licensing constraints and enable better product-specific customization.

Will these models be available to third-party developers?

While currently focused on internal integration, Microsoft plans MAI-Voice-2-Flash as standalone Azure AI service in Q1 2025. Enterprise partners may gain early access.

What's the energy impact of these new models?

Architecture optimizations reduced energy consumption by 37% for MAI-Image-2.5-Pro and 52% for MAI-Voice-2-Flash versus comparable models at equal performance levels.