Mustafa Suleyman's Microsoft AI team ships MAI-Thinking-1, the first in-house reasoning model trained from scratch on commercially licensed data with no distillation from OpenAI's GPT series. It scores 97% on AIME 2025.
Microsoft now has a frontier reasoning model it built itself. 35B active parameters, ~1T total in sparse MoE, 256K context. 97% AIME 2025, 94.5% AIME 2026.
This is the strategic break, not yesterday's Project Polaris coding model. MAI-Thinking-1 is frontier reasoning with no OpenAI lineage. Microsoft Foundry runs it in private preview today. MAI-Code-1-Flash also rolled out to GitHub Copilot users in VS Code - Microsoft says it beats Claude Haiku 4.5 on coding price-performance. Combined with the April amendment ending Azure exclusivity for OpenAI, both sides got what they wanted: Microsoft has its own models, OpenAI has its own clouds.
For PMs, the 'Microsoft = OpenAI reseller' mental model is dead. Expect more Microsoft-native models across products. Copilot pricing pressure goes UP for OpenAI as Microsoft can credibly switch defaults.
⚡ Why this matters
- First frontier-grade reasoning model from a Big Three lab that's not OpenAI / Anthropic / Google.
- 'No OpenAI data inheritance' is the moat - Microsoft owns the full lineage.
- Validates Mustafa Suleyman's hire (ex-Inflection, March 2024) as Microsoft AI head.
🔍 What happened
- MAI-Thinking-1: 35B active parameters, ~1T total in sparse MoE architecture, 256K context window.
- 97.0% on AIME 2025, 94.5% on AIME 2026 (math + multi-step reasoning benchmarks).
- Trained ground-up on commercially licensed enterprise data.
- No distillation from third-party models including OpenAI's GPT series.
- Available in private preview through Microsoft Foundry.
- MAI-Code-1-Flash: 5B-parameter coding model, rolling out to GitHub Copilot individual users in VS Code.
- Microsoft claims MAI-Code-1-Flash outperforms Claude Haiku 4.5 in coding price-performance.
- Announced June 2 at Microsoft Build 2026, Fort Mason, San Francisco.
💬 Smart takes
- Mustafa Suleyman (Microsoft AI CEO): Microsoft's strategy is to build proprietary AI alongside its OpenAI partnership.
- Microsoft: MAI-Code-1-Flash was built end-to-end by Microsoft using 'clean and appropriately licensed data.'
- Skeptic: Microsoft's prior MAI models (MAI-1 in 2024) underwhelmed at launch. Private preview benchmarks need independent verification.
🧭 Where this goes
- Within 90 days, MAI-Thinking-1 powers a Copilot Pro tier as an alternative to GPT-5.5.
- Microsoft enterprise tiers add a 'MAI-only' SKU with data lineage guarantees.
- OpenAI's pricing leverage with Microsoft drops as Microsoft has a credible internal alternative.
- EU and government customers prefer MAI for sovereignty and data lineage reasons.
🎯 Implication
- For PMs: If you ship on Azure AI Foundry, MAI-Thinking-1 is now a default option to test.
- For execs: Copilot economics shift. The 'OpenAI tax' inside Microsoft products now has an internal alternative.