Microsoft launches in-house MAI-Thinking-1 reasoning model at Build 2026

Microsoft introduced MAI-Thinking-1 at its Build developer conference, the first reasoning model built by its in-house Microsoft AI Superintelligence Team rather than sourced from OpenAI. Microsoft described it as a sparse mixture-of-experts model with about 1 trillion total parameters, of which roughly 35 billion activate per query, and a 256,000-token context window. The company said it was trained from scratch on commercially licensed data “with zero distillation” from any third-party model, OpenAI’s included — an explicit claim of independence from the partner whose models have powered Copilot since 2023.

Microsoft reported 97% on AIME 2025 and said the model matched Claude Opus 4.6 on SWE-Bench Pro, with blind side-by-side evaluations by human raters at Surge preferring it to Claude Sonnet 4.6 on overall quality. These are vendor-reported figures; no independent replication was available at launch.

MAI-Thinking-1 was one of seven MAI models Microsoft announced at Build, alongside an image model, a fast lightweight coder, and speech transcription and generation models, several of which shipped directly into GitHub Copilot and VS Code. The launch reads as a hedge: Microsoft remains OpenAI’s largest investor and infrastructure partner, but building a competitive reasoning model of its own gives it leverage and a fallback should that relationship’s terms — renegotiated only months earlier — shift again.

On 12 August 2026 Microsoft moved the model into public preview on its Foundry developer platform — its first broad availability beyond the Build unveiling. Microsoft added a reported 94.5% on AIME 2026 and said human raters preferred MAI-Thinking-1 across a 1,276-task blind evaluation; the figures remain vendor-reported and were not independently replicated.