Microsoft MAI models beat general frontier models at lower cost inside Copilot and Excel

微软MAI模型：以更低成本实现前沿能力规模化

Satya Nadella says MAI models aren't about benchmark scores—they beat general frontier models inside GitHub Copilot and Excel using fewer tokens, by learning from real product feedback through a model-agnostic evaluation system. The same template will be available to enterprise customers via Foundry. The post doesn't disclose specific performance numbers or cost comparisons.

Why it matters: Microsoft CEO lays out MAI strategy: not chasing benchmarks, but training on real feedback inside Copilot and Excel, matching frontier models with fewer tokens. Fresh angle with concrete mechanism, but the post doesn't disclose performance numbers or cost comparisons, so it st...