At Microsoft Build 2026, the company unveiled MAI-Thinking-1, its first in-house reasoning model and a clear signal of intent to reduce reliance on OpenAI for frontier capabilities. The model is a mid-sized sparse Mixture of Experts architecture with 35 billion active parameters and a 256,000-token context window, trained entirely on commercially licensed data with no distillation from any third-party model.
On AIME 2025 — a benchmark for mathematical and multi-step scientific reasoning — MAI-Thinking-1 reaches 97.0 percent, and 94.5 percent on AIME 2026. Microsoft says the model matches Claude Opus 4.6 on SWE-Bench Pro and was preferred over Claude Sonnet 4.6 in blind side-by-side evaluations conducted by Surge, Microsoft's independent human rating partner.
The launch reframes Microsoft's AI strategy. After years of leaning almost entirely on OpenAI's GPT family, Microsoft is now building its own reasoning-class models trained on clean, licensed data — a stance designed both to address enterprise IP concerns and to give Microsoft an internal alternative for high-stakes agentic workloads. The model is engineered for multi-step reasoning, software engineering tasks, and research synthesis, the same workloads enterprise customers are using GPT-5.5 and Claude for today.
MAI-Thinking-1 is available in private preview through Microsoft Foundry. Microsoft has signaled that it will expand access in stages as it benchmarks the model against external workloads and gathers safety telemetry. The release lands alongside MAI-Code-1-Flash, Microsoft's first inaugural in-house coding model, suggesting a coordinated push to build a vertically integrated MAI model family across reasoning and code.
For developers, the immediate question is whether MAI-Thinking-1 can match third-party models on real, messy production workloads. The benchmark numbers are competitive, but the private-preview gating means broader evaluation is still weeks away.
Source: [TechTimes](https://www.techtimes.com/articles/317631/20260602/microsoft-build-2026-mai-thinking-1-first-house-reasoning-model-trained-without-openai-data.htm)