MAI-Thinking-1
reasoningMicrosoft•Released on 2026-06-02
Microsoft's first in-house flagship reasoning model, unveiled at Build 2026. A ~35B active-parameter sparse Mixture-of-Experts model trained on commercially licensed data (Microsoft states it was trained without OpenAI data), with a 256K-token context window, function calling, and developer instruction support. Microsoft reports 97.0% on AIME 2025 and 94.5% on AIME 2026, and says it matches Claude Opus 4.6 on SWE-Bench Pro while being preferred over Claude Sonnet 4.6 in blind side-by-side evaluations run by its human-rating partner Surge. Available in private preview through Microsoft Foundry, with availability announced for OpenRouter, Fireworks AI, and Baseten. Public pricing is not yet finalized, and the benchmark claims have not yet been independently reproduced.
Voice of the community
“MAI-Thinking-1 reaches 97.0% on AIME 2025 and 94.5% on AIME 2026, and on SWE-Bench Pro Microsoft says it matches Claude Opus 4.6 on coding tasks.”
“Microsoft has published a preprint describing the evaluation methodology, but full reproduction of the results by independent labs has not yet occurred, leaving those benchmark claims open to challenge until confirmed externally.”
Core Specs
Scenario Scores
Pros & Cons
Pros
- +Strong reported math reasoning (AIME 2025 97.0%, AIME 2026 94.5%)
- +Microsoft says it matches Claude Opus 4.6 on SWE-Bench Pro for its weight class
- +256K-token context window with function calling and developer instructions
- +Trained on commercially licensed data — lower IP/provenance risk
- +Coming to third-party inference (OpenRouter, Fireworks, Baseten)
Cons
- −Private preview only at launch — limited access via Microsoft Foundry
- −Public pricing not yet finalized
- −Benchmark claims not yet independently reproduced (preprint methodology only)
- −Text-only at launch (no native multimodal input)
- −No community track record yet