Microsoft's MAI-Thinking-1 scores 97% on AIME 2025 with no OpenAI distillation
Microsoft's MAI-Thinking-1 — a sparse mixture-of-experts model with 35 billion active parameters out of roughly one trillion total — scores 97.0% on AIME 2025 and 94.5% on AIME 2026, trained on commercially licensed data with no OpenAI distillation, and was preferred over Claude Sonnet 4.6 by independent raters at Surge. The model enters private preview on Microsoft Foundry, giving enterprise developers on Azure a non-OpenAI reasoning option — a concrete sign that Microsoft is building its own frontier AI capability rather than reselling OpenAI's.
Source: thurrott.com ↗
MAI-Thinking-1 is the company's first reasoning model designed to offer high efficiency at a low-token cost.
Microsoft
Why this matters
- → Microsoft ships independent reasoning model, reducing dependence on OpenAI
- → 97% AIME score demonstrates competitive frontier AI capability in-house
- → Enterprise developers gain non-OpenAI reasoning option on Azure platform
Microsoft's AI independence