A New Chapter for Microsoft AI
For years, Microsoft's AI strategy was synonymous with its OpenAI partnership. That relationship remains, but Build 2026 marked a clear inflection point: Microsoft is now building its own frontier AI models, at scale, from scratch.
"Our goal is Humanist Superintelligence—AI designed to serve people, not replace them," said Mustafa Suleyman, Microsoft AI CEO. The seven models announced today form the first concrete step toward that mission, covering every major modality that matters for enterprise AI: reasoning, coding, image, voice, and transcription.
The Models in Detail
MAI-Thinking-1 — Flagship Reasoning Model
- Architecture: 35B active parameters, ~1T total parameters, sparse Mixture-of-Experts
- Performance: 97% on AIME 2025; 53% on SWE-Bench Pro (competitive with Claude Opus 4.6)
- Context window: 256K tokens (~600-page document)
- Human preference: Rated higher overall quality than Claude Sonnet 4.6 in blind evaluations across 1,276 tasks by Surge professional raters
- Availability: Private preview on Microsoft Foundry today
MAI-Code-1-Flash — Efficient Agentic Coding Model
- Size: 5B active parameters (Haiku-comparable, lower cost)
- Integration: Natively built for GitHub Copilot, VS Code, and the Microsoft stack
- Efficiency: Adaptive solution length control — concise for simple tasks, deeper for complex ones; up to 60% fewer tokens per problem
- Availability: Rolling out to VS Code Copilot individual users now (no setup required)
Full MAI Model Family
| Model | Domain | Key Highlight |
|---|---|---|
| MAI-Thinking-1 | Reasoning | Matches Opus 4.6 on SWE-Bench Pro |
| MAI-Code-1-Flash | Coding | Built into GitHub Copilot |
| MAI-Image-2.5 | Image gen & edit | Top-tier Arena score |
| MAI-Image-2.5-Flash | Image (efficient) | Ultra-fast, low-cost variant |
| MAI-Transcribe-1.5 | Speech-to-text | 5x faster than rivals, 43 languages |
| MAI-Voice-2 | Text-to-speech | 15 languages, voice cloning from short sample |
| MAI-Voice-2-Flash | Voice (efficient) | Coming soon, ultra-efficient |
MAI models are available through Microsoft Foundry with enterprise-grade security, compliance, and function calling. They also appear on OpenRouter, Fireworks, and Baseten—and for the first time, developers can fine-tune the model weights directly.
Frontier Tuning: Your Data, Your Model
Beyond the model releases, Microsoft introduced Frontier Tuning—using reinforcement learning in real-world environments to specialize MAI models on a company's own workflows. The result is a custom model that learns from actual work traces, not synthetic data.
Results so far are striking: a MAI model tuned specifically for Excel matched GPT-5.4 while being up to 10× more cost-efficient. An enterprise customer tuning for their specific standards achieved the highest win rate of any model tested at roughly 10× lower cost.
Microsoft and Mayo Clinic are jointly building a frontier healthcare AI model combining Mayo's de-identified clinical data and longitudinal insights with Microsoft's AI capabilities. The model will be owned by Mayo Clinic, deployed first within their system, and made available through Microsoft Foundry once validated. Target applications include earlier and more accurate diagnosis and treatment planning.
Built on Microsoft's Own Silicon
Microsoft is co-designing MAI models with its Maia 200 AI chip, already achieving a 1.4× efficiency boost versus comparable Nvidia hardware. The next-generation GB200 cluster at Microsoft AI is now operational, with a significant compute scale-up planned over the next 12 months.
The key principle: no shortcuts. MAI trains from scratch with clean, traceable, appropriately licensed data—no distillation from OpenAI, Anthropic, or anyone else. Full technical and safety reports are published alongside each model release.
- Microsoft's MAI family is its first fully in-house AI lineup—no third-party distillation, clean data lineage throughout
- MAI-Thinking-1 achieves Claude Opus 4.6-level coding benchmarks at medium model size (35B active parameters)
- MAI-Code-1-Flash is live in GitHub Copilot today for individual VS Code users—no setup required
- Frontier Tuning enables enterprise-specific models at up to 10× lower cost than general frontier models
- Co-designed with Maia 200 silicon; Microsoft is building long-term AI infrastructure independence