Microsoft's MAI-Thinking-1 puts a frontier reasoning model in its own stack

Microsoft released MAI-Thinking-1, its first frontier reasoning model trained in house, and put it directly alongside Claude and OpenAI's reasoning lineup. The architecture is a sparse Mixture of Experts with 35 billion active parameters out of roughly 1 trillion total, a 256k token context window, and function calling. Microsoft claims 97.0% on AIME 2025 and 94.5% on AIME 2026, and says the model matches Claude Opus 4.6 on SWE-Bench Pro for software engineering tasks.

The more telling number is from human evaluation. Across 1,276 tasks in blind side-by-side comparisons, raters preferred MAI-Thinking-1 over Sonnet 4.6. Microsoft also makes a point of saying the model was trained on "clean and appropriately licensed data" with AI-generated content excluded from pre-training, and without third-party model distillation. That framing reads as a quiet jab at competitors and a hedge against the copyright lawsuits still hanging over the rest of the industry.

The model is in private preview on Microsoft Foundry, with public access coming through MAI Playground. Microsoft did not disclose pricing, which is the variable that will decide how much of its own Copilot stack moves off OpenAI infrastructure.

Why it matters

If you build on Azure or Copilot, Microsoft now has a competitive reasoning model of its own, which gives it leverage in pricing and routing inside its stack. Watch how aggressively Copilot starts to default to MAI-Thinking-1 over GPT-class models in the coming releases.