Microsoft AI released MAI-Thinking-1, a reasoning model the company says it built without distilling from any outside lab, positioning it as the third piece of an in-house model lineup that is steadily replacing the OpenAI and Anthropic systems Microsoft has paid to run inside its own products. The pattern matters more than any single benchmark: a coding model, an image model, and now a reasoning model, each aimed at cutting the per-query cost Microsoft otherwise pays a partner lab. Microsoft frames the release inside a program it calls the Hill-Climbing Machine, a training pipeline meant to compound gains from better data, rewards, and compute rather than borrow capability from a rival.
MAI-Thinking-1 is a sparse mixture-of-experts model with 35 billion active parameters and roughly 1 trillion total parameters, a footprint Microsoft describes as smaller than the largest frontier systems. On SWE-Bench Pro, a benchmark for software engineering tasks, Microsoft says the model performs on par with Anthropic’s Claude Opus 4.6. On the AIME 2025 and AIME 2026 math competitions, Microsoft reports scores of 97.0 percent and 94.5 percent. In head-to-head evaluations judged by human raters, an exercise Microsoft designed and ran itself, the company says its outputs were preferred over Anthropic’s Sonnet 4.6.
Every one of those figures comes from Microsoft, measured against comparisons Microsoft chose. The release does not cite a third-party evaluator or an independently run benchmark. AI Insiders reported yesterday on Microsoft’s coding model, MAI-Code-1.1-Flash, where outside testing painted a less generous picture than the company’s own launch materials did. The same gap between self-reported and independently verified performance applies here until someone outside Redmond runs the numbers.
The strategic logic is about margin, not raw capability. Microsoft has committed billions of dollars to OpenAI compute deals and licenses Anthropic’s models for parts of Copilot and GitHub. Every query Microsoft can route to a model it trained and owns outright is a query it no longer pays another lab’s API rate to serve. A medium-sized reasoning model with a 35-billion-parameter active footprint fits that math well: it is cheap enough to run at the volume Copilot and enterprise customers generate, without the inference cost of a much larger frontier system.
Microsoft also stresses that the model learned its capabilities directly rather than imitating outputs from GPT or Claude during training. The company’s argument is that models taught by copying a rival inherit that rival’s design limits and adapt poorly to new situations, while a model trained on data Microsoft can trace and audit is easier to steer and correct. That is a defensible engineering position, and it doubles as a hedge against the licensing and intellectual-property disputes that have followed distillation-heavy training approaches across the industry this year.
None of this settles whether MAI-Thinking-1 is actually stronger than what Microsoft already licenses from its partners. It confirms that Microsoft now has three homegrown model lines improving on a predictable cadence, each aimed at workloads it currently pays OpenAI and Anthropic to serve. Enterprise buyers running coding, math, or knowledge workloads through Copilot should watch which model gets routed to their queries by default over the next two quarters. That default, not a blind evaluation Microsoft conducted on its own, is where the margin argument gets tested against real usage.
Microsoft AI announced MAI-Thinking-1 on its company blog in August 2026.
