微软正在多个 Copilot 产品（包括 Excel 和 Outlook）中，用自研模型取代来自 OpenAI 和 Anthropic 的 AI 模型。
据彭博社报道，微软自研的 MAI 模型已在 Excel 和 Outlook 中每周处理数万次请求。此前，这两款应用更依赖 OpenAI 和 Anthropic 的模型。
自研模型目前处理的请求量仍只占总请求的一小部分，但微软希望持续削减在第三方 AI 上的支出。MAI 模型也已应用于 GitHub Copilot，一款自研的转录模型预计很快将在 Teams 中推出。
在 Build 大会上，微软发布了七款新 AI 模型，其中包括其首款推理模型 MAI-Thinking 1。微软声称，根据人工评估，该模型在编程能力上可与 Sonnet 4.6 和 Opus 4.6 媲美。但当时公布的基准测试结果却显示，Thinking-1 与 OpenAI 和 Anthropic 的竞品差距明显，大致与 Deepseek V3.2 处于同一水平。
用户可能以相同价格获得性能较弱的 AI
对于 Copilot 和 Office 用户而言，这可能意味着以相同价格获得性能较弱的 AI，以便微软降低自身成本。微软 AI 负责人穆斯塔法·苏莱曼在六月公开承认了这一计划：“我们向 Anthropic 支付了巨额费用——因此我们的目标是减少并最终消除这笔成本。”而正是这家微软，此前刚刚主张与 OpenAI 和 Anthropic 的供应商锁定是件坏事，并表示希望成为平台中立的替代方案。
微软 CEO 萨提亚·纳德拉还暗示，AI 计费可能进一步转向基于使用量的定价，而非固定费率订阅。一种可能的方案是，将更便宜的 MAI 模型设为默认选项，而来自 OpenAI 或 Anthropic 的第三方模型则作为高级附加功能，需额外付费。这样一来，微软就能将其在 OpenAI 和 Anthropic 上的开支以附加费的形式转嫁给用户。
微软还声称，MAI 模型是在干净、获得商业许可的数据上训练的，因此对企业使用是安全的。但根据技术论文，微软使用了 Common Crawl 数据集——这是一个可自由访问的网络数据集合，将其用于 AI 训练在法律上尚无定论。其他所有 AI 公司也都这么做，但微软却将自己使用的训练数据描绘得格外干净。
Microsoft is replacing AI models from OpenAI and Anthropic with its own in-house models across several Copilot products, including Excel and Outlook.
Microsoft's in-house MAI models are already processing tens of thousands of requests per week in Excel and Outlook, Bloomberg reports. Both apps previously relied more heavily on models from OpenAI and Anthropic.
The in-house models still handle only a small fraction of total requests, but Microsoft wants to keep cutting its spending on third-party AI over time. The MAI models are also available in GitHub Copilot, and a proprietary transcription model is expected to ship in Teams soon.
At the Build conference, Microsoft unveiled seven new AI models, including MAI-Thinking 1, its first reasoning model. Microsoft claimed it could match Sonnet 4.6 and Opus 4.6 in coding based on human evaluations. But the benchmarks released at the time told a different story, with Thinking-1 trailing the competition from OpenAI and Anthropic by a wide margin and landing roughly on par with Deepseek V3.2.
Customers may get less capable AI for the same price
For Copilot and Office customers, this could mean paying the same amount for weaker AI so that Microsoft can lower its own costs. Microsoft's head of AI, Mustafa Suleyman, openly acknowledged this plan in June: "We pay a lot of money to Anthropic—so our goal is to reduce and ultimately eliminate that cost." This is the same Microsoft that just argued vendor lock-in with OpenAI and Anthropic is a bad thing and that it wants to be a platform-neutral alternative.
Microsoft CEO Satya Nadella also hinted that AI billing could shift further toward usage-based pricing instead of flat-rate subscriptions. One possible setup would make cheaper MAI models the default, with third-party models from OpenAI or Anthropic available as premium add-ons at extra cost. That way, Microsoft would pass its OpenAI and Anthropic expenses on to customers with a surcharge.
Microsoft also claimed that the MAI models are trained on clean, commercially licensed data, making them safe for businesses. But according to the technical paper, Microsoft used the Common Crawl dataset, a collection of freely accessible web data whose use for AI training is not legally settled. Every other AI company does the same thing, but Microsoft portrays its own training data as particularly clean.