Published: June 03, 2026, 17:09 UTC
Microsoft released its first advanced reasoning AI model at Build 2026 on Tuesday, marking a major step in the company’s drive to reduce its dependence on OpenAI’s technology. MAI-Thinking-1 is a sparse mixture-of-experts model with 35 billion active parameters out of approximately 1 trillion total — and it was trained entirely from scratch on clean, commercially licensed data without any distillation from third-party models, including OpenAI’s ([Microsoft AI](https://microsoft.ai/news/introducing-mai-thinking-1/)).
A model designed for reasoning. MAI-Thinking-1 uses chain-of-thought reasoning internally before producing final outputs, targeting complex multi-step problems across mathematics, science, and software engineering. Mustafa Suleyman, CEO of Microsoft AI, described it as a general-purpose reasoning engine rather than a narrow benchmark specialist ([The Verge](https://www.theverge.com/tech/941664/microsoft-ai-model-reasoning-mai-thinking-1-build-2026)).
The model scores 97.0% on AIME 2025 (a challenging mathematics benchmark) and 94.5% on AIME 2026, demonstrating strong mathematical reasoning. On SWE-Bench Pro, the software engineering benchmark, Microsoft claims MAI-Thinking-1 goes “toe-to-toe with Claude Opus 4.6” — though independent analysis places it at roughly 52.8%, below GPT-5.4 (57.7%), Kimi K2.6 (58.6%), and GLM-5.1 (58.4%). In blind human side-by-side evaluations, it was preferred over Claude Sonnet 4.6 ([Microsoft AI](https://microsoft.ai/news/introducing-mai-thinking-1/)).
A 256,000-token context window (~600 pages of text) and support for the Chat Completions API make the model accessible to developers through standard tooling. Training used 8,192 NVIDIA GB200 GPUs on a pre-training corpus of approximately 30 trillion tokens.
Building independence from OpenAI. The launch carries unusual strategic weight. Microsoft was contractually barred from independently pursuing advanced AI until October 2025 under its partnership agreement with OpenAI. That restriction has expired, and MAI-Thinking-1 is the first major fruit of Suleyman’s accelerated in-house push since joining Microsoft to lead its consumer AI division.
Microsoft is making a deliberate point of what it calls “clean data provenance” — the model was trained without using AI-generated content in its pre-training data and without distilling outputs from third-party models. The company markets this as an enterprise differentiator: auditable training data with no intellectual-property contamination risk.
The Hill-Climbing Machine. MAI-Thinking-1 is part of a broader framework Microsoft calls the “Hill-Climbing Machine,” a co-designed pipeline meant to make every component of model development improvable over time — data, rewards, evaluation environments, and compute. The company positions it as a continuous improvement platform rather than a one-shot release.
Pricing and availability. The model is in private preview via Azure AI Foundry. Third-party providers Fireworks AI, Baseten, and OpenRouter offer it at launch — a deliberate anti-vendor-lock-in strategy targeting enterprise developers who deliberately avoid Azure. First-party pricing has not been finalized, but for reference, Microsoft’s MAI-Code-1-Flash is priced at $0.75 per million input tokens and $4.50 per million output tokens.
The seven-model family. MAI-Thinking-1 launches alongside six other models at Build 2026: MAI-Code-1-Flash (a 5-billion-parameter agentic coding model shipping today to GitHub Copilot and VS Code, scoring 51% on SWE-Bench Pro), MAI-Image-2.5 (image generation and editing), MAI-Voice-2 (multilingual text-to-speech in 15 languages), and MAI-Transcribe-1.5 (speech recognition).
Where it falls short. Terminal-Bench is a notable weakness — Microsoft concedes the model was not trained on terminal-interaction environments, limiting its effectiveness for autonomous agent use cases. The model also does not lead the field on any single benchmark; it is competitive across the board but rarely state-of-the-art. Consumer access is not yet available; the launch is developer- and enterprise-only.
Still, as the first output of Microsoft’s post-restriction in-house reasoning effort, MAI-Thinking-1 signals that the company intends to be more than a distribution partner for other people’s models. Whether it can close the gap to GPT and Claude on the next release cycle will tell whether the Hill-Climbing Machine is a genuine platform or just a new name for catching up.
Sources: [Microsoft AI — Introducing MAI-Thinking-1](https://microsoft.ai/news/introducing-mai-thinking-1/) (June 2, 2026); [The Verge — Microsoft’s first advanced reasoning AI](https://www.theverge.com/tech/941664/microsoft-ai-model-reasoning-mai-thinking-1-build-2026) (June 2, 2026); [Hacker News discussion](https://news.ycombinator.com/item?id=48374528) (June 2, 2026)