微软称新 AI 编程模型以 5% 成本媲美 Claude Opus 5.5

微软在 Ignite 2026 上发布了三款新的 AI 模型。
该公司声称,在仅用 5% 成本的情况下,达到了 Claude Opus 96% 的性能。
更便宜的模型可能为微软节省数亿美元。
在今天的 Microsoft Ignite 2026 主题演讲中,该公司推出了两款新的前沿 AI 模型,专为公司所称的“日常”编程任务打造。但更大的看点在于:这些模型对微软的 AI 成本及其对 OpenAI 的依赖意味着什么。
微软声称,其 MAI Code 2 模型展现出 Claude Opus 5.5 96% 的性能,而价格仅为 Opus 5.5 的 5%。
微软表示,MAI Code 2 的性能可与 GPT-6.1 Sol 或 Opus 5.5 相媲美,而成本仅为后者的一小部分。微软的 Code 2 Flash 模型可与较低端的 GPT-6 Luna 和 Claude Haiku 5.5 相比,成本也低得多。
这并不意味着我们这些每月支付 100 美元订阅费的人会突然以五美元获得同等的服务。相反,这些是基于 API 用量的指标。
微软自研的 AI 模型可能减少其对 OpenAI 和 Anthropic 的依赖。
在 Satya Nadella 的整场主题演讲中,Microsoft 展示的幻灯片提到了多种基础模型,这些模型旨在与 Copilot 及 Microsoft 的其他产品协同工作。众所周知,Microsoft 对 OpenAI 有着数十亿美元的投资,这可以追溯到 2023 年,并且该公司围绕 OpenAI 的知识产权构建了许多产品和服务。
但如今,微软正把其自研模型与 Anthropic 和 OpenAI 的产品放在一起做基准测试,同时还宣称正将这些新模型作为可选项集成到 GitHub Copilot、Microsoft 365 和 Fabric 中。
Nadella 在今天的主旨演讲中明确传达的一点是,AI 正日益成为 Microsoft 所做一切的核心。此前,几乎所有这些 AI 操作都意味着公司要向那些前沿模型供应商支付一定的 Token 费用,或者由客户支付同样的费用。
但有了自家性能强劲的模型,Microsoft 对外部模型供应商的依赖可能会大幅降低,尤其是考虑到该公司庞大的产品线中会产生数量惊人的、相当基础且常规的 AI 请求。
Microsoft 把其编码模型变成了面向 Office 的 AI 员工
回顾 ChatGPT Work 和 Claude Cowork 这类智能体式工作模型的发展历程,你或许还记得它们最初都是智能体式编码工具。当用户把 Claude Code 和 OpenAI Codex 中的编码框架用于超出原本用途的生产力应用时,这些 AI 公司注意到了这一现象,并打包推出了更面向非程序员、以工作为核心的智能体式工具。
编码智能体与工作智能体之间有很多相似之处,包括理解复杂环境、使用工具改变事物并产出可靠结果。事实证明,编码是开发和测试这些能力的极佳环境,尤其因为代码输出具有高度可衡量性。编码智能体是训练非编码工作智能体的有效训练场。
这在今天的公告中变得非常有趣。除了 Code 2 和 Code 2 Flash,Microsoft 还发布了 MAI Work Flash,并称其衍生自 Code 2 Flash。Work Flash 可以帮助在 Excel 中生成公式、处理 Word 中的文档和指令、跟踪决策并从 Teams 中提取有用信息,等等。
Microsoft 最大的 AI 模型客户可能就是 Microsoft 自己
想想这些任务对 AI 模型的智能水平要求有多高。其实并不高。只要模型足够好,它们就够用了。
如果 Microsoft 能够降低开支,并以低得多的价格使用可靠的模型,那么该公司自身将能够在整个软件经济中节省相当多的资金,同时保持更深层次的垂直整合,从而保持控制力。
让我们来算一些假设性的数字。假设一家公司每次 AI 请求花费 0.01 美元,而有了这些新模型,这一成本可以降到每次请求 0.001 美元。
一家每日推理成本为 100 万美元的公司(例如 Microsoft 这样的公司)可以将每日成本降至约 10 万美元。按一年累计,更换模型每年可节省 3.25 亿美元。
这里一百万,那里一百万。很快,你谈的就是真金白银了。
没有直接证据表明微软正试图为前沿模型厂商的产品打造替代模型。但如果把目光投向微软为其所售软件开发自有引擎这件事,这确实是一个有趣而颇具说服力的故事。
微软的基准测试成绩令人瞩目,但重要问题仍待解答
微软表示,它公布这些数字时采用的是 50% 的首发折扣价。这就是我在上文以每次请求 0.001 美元、而非每次请求 0.0005 美元作为对比基准的原因。
基准测试的选取可能存在“挑樱桃”之嫌。这些数字确实令人瞩目。但众所周知,世上有谎言、该死的谎言,还有基准测试结果。我相当确信微软没有在这些数字上注水太多,因为相比它自己的客户,微软从这些新模型中获益更多,但即便如此,这仍是个悬而未决的问题。
该公司也明确表示,这些新模型并非为处理最具挑战性的智能体式任务和研究类工作负载而设计。你大概不会用它们去攻克一道悬置数十年的数学难题,但你很可能会用它们来重新排版你的 PowerPoint。
说到重新排版 PowerPoint,微软还强调,这些新模型在可视化任务上表现格外优异。这不仅非常适合应对 UI 设计挑战,也非常适合图形与幻灯片制作。
资深特约编辑
计算机科学教授出身的人工智能创新者 David Gewirtz 已在人工智能前沿推动进步超过三十年,并获得了 Sigma Xi 工程研究奖。他撰写了一篇开创性分析,帮助塑造了早期围绕 AI 伦理的讨论。他还是 AI 产品商业化的先驱,包括 AI 语言和基于知识的系统。他是 Al Editor 的设计者,这是一个用于解析、分类和识别新闻报道的实验性人工智能引擎。David 是人工智能促进协会(Association for the Advancement of Artificial Intelligence)的成员。他目前任职于 InfraGard 的人工智能威胁与缓解跨部门委员会(AI CSC),InfraGard 是 FBI 与行业领袖之间的合作组织,旨在保护美国关键基础设施。他著有《Where Have All the Emails Gone?》,以及《How To Save Jobs》和《The Flexible Enterprise》。查看完整简介
译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。
查看英文原文(Microsoft says its new AI coding model rivals Claude Opus 5.5 at 5% of the cost)
Microsoft unveiled three new AI models at Ignite 2026.
It claims 96% of Claude Opus performance at 5% of the cost.
Cheaper models could save Microsoft hundreds of millions.
At today’s Microsoft Ignite 2026 keynote, the company introduced two new frontier AI models built for what the company describes as “everyday” coding tasks. But there’s a bigger story: what these models might mean for Microsoft’s AI costs and its dependence on OpenAI.
Microsoft claims its MAI Code 2 model shows 96% of Claude Opus 5.5 performance at 5% of Opus 5.5’s price.
Microsoft says MAI Code 2 offers comparable performance to GPT-6.1 Sol or Opus 5.5 at a fraction of the cost. Microsoft’s Code 2 Flash model is comparable to the lower-end GPT-6 Luna and Claude Haiku 5.5, also at considerably lower cost.
This doesn’t mean that those of us who pay for $100/month subscriptions will suddenly get the equivalent for five bucks. Instead, these are usage-based metrics based on API usage.
Microsoft home-grown AI models could reduce its reliance on OpenAI and Anthropic
Throughout Satya Nadella’s keynote, Microsoft showed slides referencing a wide range of foundation models that are intended to work with Copilot and other Microsoft offerings. Microsoft famously has a multi-billion-dollar investment in OpenAI dating back to 2023, and has built many of its products and services around OpenAI intellectual property.
But now, Redmond is benchmarking its internally built models against Anthropic and OpenAI offerings, and also touting that it’s integrating these new models as optional choices for GitHub Copilot, Microsoft 365, and Fabric.
One thing that Nadella made clear in today’s keynote is how central AI is becoming in everything Microsoft does. Nearly all of those AI operations previously involved the company paying some token fees back to the frontier model providers or customers doing the same thing.
But with high-performing models of its own, Microsoft may become substantially less dependent on outside model providers, especially for the incredible volume of fairly basic and routine AI requests that occur across the company’s vast product line.
Microsoft turned its coding model into an AI worker for Office
If you think back over the history of agentic work models like ChatGPT Work and Claude Cowork, you may recall that they originated as agentic coding tools. As users adopted the coding harnesses in Claude Code and OpenAI Codex to off-label productivity applications, the AI companies noticed and packaged work-centric agentic tools that were more non-programmer-friendly.
There’s a lot of similarity between a coding agent and a work agent, including understanding a complex environment, using tools to change things, and producing reliable results. Coding turns out to be a very good environment for developing and testing these capabilities, particularly because code output is highly measurable. Coding agents are effective training grounds for non-coding work agents.
This becomes really interesting in today’s announcement. Along with Code 2 and Code 2 Flash, Microsoft announced MAI Work Flash, which it describes as having been derived from Code 2 Flash. Work Flash can help generate formulas in Excel, work with documents and instructions in Word, track decisions and extract useful information from Teams, and more.
Microsoft’s biggest AI model customer might be Microsoft itself
Think about the level of intelligence these tasks are demanding from the AI models. Not that much. As long as the models are good enough, they’ll be good enough.
If Microsoft can reduce its expenses and use reliable models at a much lower price, the company itself will be able to save quite a bit of money across its entire software economy, while at the same time maintaining a deeper level of vertical integration and, therefore, control.
Let’s run some hypothetical numbers. Let’s say a company is spending $0.01 per AI request, which, with these new models, can drop down to $0.001 per request.
A company with a daily inference cost of a million dollars (a company like Microsoft, for example) could drop that daily cost to about $100,000. Add that up across a year, and the change in models could save $325 million annually.
A million here, a million there. Pretty soon, you’re talking real money.
There’s no direct evidence that Microsoft is trying to produce replacement models for the frontier providers’ offerings. But it is an interesting and compelling story if we look at Microsoft developing its own engines for the software it sells.
Microsoft’s benchmark claims are impressive, but important questions remain
Microsoft says the numbers it’s reporting at are using a 50% introductory discount. That’s why I measured against $0.001 per request instead of $0.0005 per request above.
Benchmark selection could be cherry-picking. The numbers are impressive. But as we all know, there are lies, damned lies, and benchmark results. I’m fairly sure Microsoft isn’t padding these numbers too much, because it benefits from the new models more than even its own customers, but still. It’s an open issue.
The company also described clearly that these new models are not designed to handle the most challenging agentic and research workloads. You’re probably not going to use them to solve a decades-old mathematics challenge. But you’re very likely to use them to reformat your PowerPoint.
Speaking of reformatting your PowerPoint, Microsoft also emphasized that these new models did unusually well on visualization assignments. That’s a very good fit for not only UI design challenges but also graphics and slide production.
You can follow my day-to-day project updates on social media. Be sure to subscribe to my weekly update newsletter, and follow me on Twitter/X at @DavidGewirtz, on Facebook at Facebook.com/DavidGewirtz, on Instagram at Instagram.com/DavidGewirtz, on Bluesky at @DavidGewirtz.com, and on YouTube at YouTube.com/DavidGewirtzTV.
Please accept cookies to view the comments powered by Disqus.
Senior Contributing Editor
Computer science professor turned AI innovator David Gewirtz has spent over thirty years driving advancements at the forefront of artificial intelligence and is the recipient of the Sigma Xi Research Award in Engineering. He wrote a pioneering analysis that helped shape the early conversation around AI ethics. He was also a pioneer in the commercializing of AI products, including AI languages and knowledge-based systems. He is the designer of the Al Editor, an experimental artificial intelligence engine for parsing, classifying, and identifying news stories. David is a member of the Association for the Advancement of Artificial Intelligence. He currently serves on the Artificial Intelligence Threat and Mitigation Cross-Sector Council (AI CSC) for InfraGard, a partnership between the FBI and industry leaders for the protection of U.S. Critical Infrastructure. He is the author of Where Have All the Emails Gone?, as well as How To Save Jobs and The Flexible Enterprise. See full bio
原文为英文,正文由自动抓取并保留原文。
原文发布于 ZDNet(Microsoft says its new AI coding model rivals Claude Opus 5.5 at 5% of the cost),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。