Microsoft Ignite 发布 MAI Code 2 等自研模型,Autopilot 常驻智能体开始推送
微软这次是真发了自研模型:MAI Code 2 号称用 5% 成本跑到 Claude Opus 5.5 的 96% 性能,直接在 GitHub Copilot 里能用。Autopilot 常驻智能体也开始推送,走的是和 OpenAI Dots 类似的单智能体路线。关心企业 token 成本和 Copilot 生态的人都该看看这篇一手观察。
Microsoft Ignite 上,微软公布了多个自研 MAI 模型:MAI Code 2 和 Code 2 Flash 已上线 GitHub Copilot,官方称 MAI Code 2 以 5% 的成本达到 Claude Opus 5.5 性能的 96%。另有 MAI Image 2.7、语音到语音模型 MAI Realtime 1,以及小型路由模型 Microsoft-Decision-1.1。
产品层面,Autopilot 是带独立身份和记忆的常驻智能体,先面向 Frontier 早期客户推送;Copilot Home 将 Chat 与 Cowork 整合,Word、Excel、PowerPoint 进入 Copilot 应用并正式可用。Workspaces 取代 Notebooks,成为人和智能体协作同一项目的共享空间。Microsoft IQ 作为上下文层,让企业智能体读取邮件、文件和内部数据。
作者指出微软的核心押注是效率:企业 token 成本持续上涨,微软靠便宜自研模型加超越简单路由的编排(如 GitHub Fusion 中弱模型起草、强模型接手或互相审查)控制成本,同时把客户留在 Office 生态。作者也提醒,Autopilot 单智能体、单长会话的设计类似 OpenAI 的 Dots,需避免重蹈 OpenAI 覆辙——Grok Bot 等产品支持任意数量智能体和更深个性化,值得微软警惕。
A short review of Microsoft Ignite + honest take:
I spent the last few days at Microsoft Ignite. I got to watch the keynote and do two interviews, with Kyle Daigle (@kdaigle COO of GitHub and CMO Developer at Microsoft) and Sergey Fayfer ( @fayfers CVP Copilot Product). Two absolutely nice guys. On top of that, I had a Q&A with Microsoft CEO Satya Nadella in a very small group. Great as always, you just get really interesting insights.
First, a quick recap of what Microsoft announced:
- New in-house MAI models, including MAI Code 2 and Code 2 Flash (live in GitHub Copilot), MAI Image 2.7 and the speech-to-speech model MAI Realtime 1
- Autopilot, an always-on agent with its own identity and memory, starting to roll out to Frontier customers (Microsoft's early access program)
- Copilot Home, where Chat and Cowork come together, plus Word, Excel and PowerPoint inside the Copilot app, both generally available
- Workspaces, shared spaces where people and agents work on one project (the successor to Notebooks)
- Microsoft-Decision-1.1, a small router model that picks the model and effort level for each task and powers the Auto mode in Cowork
- Microsoft IQ, the context layer that gives company agents access to mails, files and data
I want to highlight two things:
1) Microsoft is betting on efficiency. Cheaper tokens matter in business. With MAI Code 2 they continue on the path GitHub already took: their own cheap models, plus orchestration that goes way beyond simple routers. GitHub's Fusion picks the models for each task and lets them work together. A cheaper model starts and a stronger one only takes over when needed, or one model drafts and another reviews. Microsoft says MAI Code 2 reaches 96% of Claude Opus 5.5's performance at 5% of the cost.
And credit where it's due: Microsoft, with its focus on business and enterprise through Office, will be successful with this. I say that out of conviction. You hear from everywhere that token costs keep rising in companies. Microsoft shows up as the provider with ideas on how to cut those costs, keep working smart and keep customers inside its own Office ecosystem at the same time.
2) And that ecosystem got a lot bigger. Autopilot, Microsoft's always-on agent for work, is now rolling out, starting with Frontier customers. That matters too. Think about how many customers Microsoft has through Office. This is a big step toward getting the working world used to personalized agents. They have the context of your company, they are tailored to each person, and they are very similar to OpenAI's Dots.
To be fair, I also said in my conversations that Microsoft has to be careful not to make the same mistake as OpenAI here. What makes Grok Bot (among others) so good is that you can create as many agents as you want, with team meetings, a chief of staff and much deeper personalization. Autopilot feels more like OpenAI's Dots: one agent, one long session, even though you can create several sessions with Autopilot. I'd like to see more personalization, but I heard Microsoft's always-on agent is still a work in progress. The same applies to Copilot. Routing is often selected by default. However, my distrust likely stems from the GPT-5 release, when routing was first truly implemented - and it worked rather poorly.
One more thought that came to me during the event: Microsoft and GitHub are going their own way more and more, with their own models. With MAI Code 2 and Fusion they get close to frontier models like Opus 5.5 at a much lower cost. To me it looks like Microsoft is looking for independence. More and more of their business runs on their own models, together with attractive offers for their customers, and that could actually become a moat. We'll see.
Overall a really interesting look behind the scenes. Microsoft is slimming down and simplifying its interface and bringing it all together, and I'm quite sure they listen to feedback. That's why I look back at Ignite in a positive way.
It was a great experience and im really looking forward to what Microsoft is going to release and update. And who knows, maybe in a few months even Microsoft is starting to compete with OpenAI, Anthropic and Google (though that's not very likely ;)
本条情报归属 AITOP「产品」核心栏目,重点聚焦北美前沿实验室与硅谷科技巨头。在跨区域产业观察坐标中,该动态直接关联「AI 开发者工具链、Coding Agent 与工程化实践」「中美大模型技术代差与商业化对照」研究主线,反映出当前技术演进与商业落地的最新风向。涉及核心实体与技术议题包括:Microsoft、MAI Code 2、GitHub Copilot、Autopilot,读者可通过下方标签进入对应专题页查阅历史脉络与关联对比。
信源核验说明:本文原始线索采编自「kimmonismus」(@kimmonismus),经 AITOP 多维评分管线结构化提炼。