微软用更便宜的自研模型替换产品中的外部 AI 模型
GPT-Image-2,提示词来自 THE DECODER
微软的 AI 部门在其 Ignite 大会上发布了七款专用模型,覆盖代码、IT 安全、图像编辑、语音和工作流决策。这些模型都旨在降低全天候运行 AI 智能体的运营成本。
这些模型以 OpenAI 和 Anthropic 的系统为对标对象,提供更便宜的替代方案。编码模型 MAI Code 2 在内部基准测试中几乎追平 GPT-6.1 Sol 和 Claude Opus 5.5,而成本仅为其一小部分。MAI Cyber 2 Flash 在 CyberGym-E2E 漏洞测试中略微胜过 GPT-6 Astra。
除了图像模型 MAI Image 2.7、语音模型 MAI Realtime 1 以及基于 Google Gemma 的路由模型 Decision 1.1 之外,微软还将允许企业客户在 Azure 云中用自己的数据微调部分模型的权重。
Microsoft AI(MAI)正以面向代码、安全、图像、语音和路由的新模型瞄准全天候运行的 AI 智能体。其主要的衡量标尺,正是自家合作伙伴 OpenAI 的模型阵容。
在其 Ignite 大会上,MAI 发布了七款新模型,一次性批量推出,而不是分散在全年陆续发布,这与它 6 月在 Build 上的做法一致。微软主要从性价比角度为这一新阵容定位。该公司提到,希望全天候运行 AI 智能体的客户,其 Token 支出正在迅速攀升。微软表示,其代码模型针对日常企业编码任务,尚未针对基础研究或长周期智能体式工作流等更困难的工作进行优化。
MAI Code 2 以促销价格逼近 GPT-6.1 Sol 和 Opus 5.5 的水平
据微软称,MAI Code 2 在四项基准测试中平均达到 GPT-6.1 Sol 性能的 98%,而成本仅为后者的 40%。与 Claude Opus 5.5 相比,它达到其 96% 的性能,成本仅为其 5%。微软在 DeepSWE 1.1、Terminal Bench 2.1、SWE-Atlas 和 FrontierCode v1.1 上进行了测试。其中三项测试在微软内部的 MSBench 平台上运行,使用 GitHub Copilot CLI,且不使用子智能体。微软为其自家模型计算了 50% 的促销折扣。
微软表示,在前端工作(如 UI、游戏、数据可视化和 SVG 图形)的盲测对比中,人类评估者在 64% 的情况下更偏好 MAI Code 2 的输出,而非 GPT-6 Sol。
MAI Code 2 在成功率上略微落后于 GPT-6.1 Sol 和 Opus 5.5,但按促销价计算,每项任务的成本仅为前者的一小部分。|图片:微软
作为 MAI-Code-1.1-Flash 的继任者,MAI Code 2 Flash 面向代码解释、定向编辑和常规修复。据说它能与 GPT-6 Luna 和 Claude Haiku 5.5 相匹敌,成本分别为后两者的 65% 和 20%。
微软从 Flash 模型衍生出 MAI Work Flash。它运行于 Microsoft 365 和 Fabric IQ 中。微软称,在内部 Excel 和 PowerPoint 基准测试中,它与 GPT-6 Luna 相当。在 Word 中,其表现与 GPT-6.1 Sol 相近,同时输入 Token 消耗约减少 42%。在 Power BI 中生成 DAX 查询方面,它优于 GPT-5.6 Sol、GPT-6 Astra 和 Opus 5.5。
在 GitHub Copilot 中,MAI Code 2 的促销价为每百万输入 Token 0.50 美元、每百万输出 Token 2.50 美元,该价格持续至 2027 年 5 月 17 日,之后为 1 美元和 5 美元。MAI Code 2 Flash 在此之前为 0.05 美元和 0.25 美元,之后涨至 0.10 美元和 0.50 美元。
安全模型,基准测试经过清理
微软表示,MAI Cyber 2 Flash 在 CyberGym-E2E 漏洞基准测试中解决了 92.3% 的任务。GPT-6 Astra 达到 91.5%,但需要两倍的推理预算。按任务计算,MAI 模型的成本不到 GPT-6 Astra 的 2%,而每 Token 成本仅为其百分之一。在 WebExploitBench 上,它得分 40.4%,领先 GPT-5.6 Cyber 的 38.4%。
在 CyberGym-E2E 上,MAI Cyber 2 Flash 解决的任务略多于 GPT-6 Astra,同时使用的 Token 更少、成本仅为后者的一小部分。| 图片来源:Microsoft
在评估过程中,微软发现约 900 项 CyberGym 任务中有 10% 所依据的开源项目和漏洞报告,与其自身训练数据的一部分相同。微软表示,参考解法从未进入训练数据。即便如此,微软仍对所有模型都移除了这些任务,仅对其余任务进行评分。
目前尚缺少与 GPT-6 Cyber 的对比。该模型仅向部分客户开放。微软表示,带工具的智能体式基准测试仍需额外的安全措施。该公司计划在未来几周公布这些结果。
其上下文窗口达 100 万 Token,是 MAI Cyber 1.1 Flash 的四倍。输出 Token 的成本比前代低 86%。新闻稿称,MAI Cyber 系列模型已承担微软安全平台 MDASH 上近 40% 的扫描工作。
微软将 MAI Cyber 2 Flash 定价为每百万输入 Token 0.10 美元、每百万输出 Token 0.50 美元,而 GPT-6 Astra 分别为 10 美元和 50 美元。在 CyberGym-E2E 上,这意味着每项任务成本为 0.45 美元,而非 34.38 美元。
微软的决策模型现在运行在 Google 的 Gemma 之上。
Microsoft-Decision-1.1 负责路由、分类、优先级排序和工作流控制。该模型从开发者批准的动作中进行选择,并返回经过校准的置信度分数。当不确定性较高时,应用可以上报升级处理,而不是依据错误决策采取行动。如果信息不完整,该模型在设计上会选择弃权。
它的基础是 Google 的混合专家模型 Gemma 4 26B A4B。其前代模型基于 Qwen3.5-9B 构建。Microsoft 使用监督微调和强化学习,在数百个合成及授权数据集上训练了这个新模型,并将其运行在专有推理引擎上。
在 Artificial Analysis 的 Decision Index v2.1 上,Decision 1.1 得到 68 分。这比其前代高出 6 分,并高于 Jev 1.14、Gemini Decision 1 和 GPT-6 Luna (Decisions)。Microsoft 还在 102 项内部基准测试、超过 385,000 个问题上对其进行了测试,在八个决策模型中取得了最高的平均准确率。
该模型现在可以接受图像输入,包括截图。在 Copilot 中,它决定一个请求是交给 Copilot 还是 Cowork 处理。Microsoft 表示,其准确率与大型语言模型相当,但速度快 2.2 倍、成本低 4.4 倍。在 Cowork 中,其 Auto 功能控制一项任务能获得多少推理投入。
Microsoft 对每百万输入 Token 收费 0.042 美元,在 EU 数据区为 0.05 美元。输出 Token 免费。
在 Artificial Analysis 的 Decision Index 上,Decision 1.1 比其前代提升 6 分,响应速度提升超过一倍。| 图片来源:Microsoft
Ignite 的 Book of News 中又出现了一款专用模型。Microsoft-Retriever-1 为 Foundry IQ 中的智能体从企业文档里汇集相关信息。该模型不是只执行一次搜索,而是会规划查询、评估结果并不断追问,直到掌握足够材料给出有充分依据的回答。微软表示,这比此前的方法更快、更高效。Retriever-1 以预览形式发布。Foundry 中的 Model Router 现在也能从扩充后的模型池中自动选择模型。
图像编辑能力刚刚略胜 OpenAI 一筹
MAI Image 2.7 以 5,478 票获得的 1,215 Elo 分数领跑 Artificial Analysis 图像编辑排行榜。它比 GPT Image 2.6 Sunburst(max)高出 4 分,而成本仅为其五分之一。在文生图生成方面,该模型在该排行榜上位列第四。MAI Image 2.7 每 1,000 张图像的成本约为 35 美元。Flash 版本约为 10 美元。
据 Artificial Analysis 称,MAI Image 2.7 和 2.7 Flash 以低得多的价格达到了与 GPT Image 2.6 Sunburst 相当的编辑质量。|图片:Microsoft
该模型可以更改单个物体的颜色、材质或位置,而不改动图像的其余部分。它还能生成透明背景。更小的 Flash 版本成本不到其三分之一,并驱动 MAI Labs 的新图像编辑器 Canvas。自 6 月以来,MAI Image 一直是 Bing Image Creator 的默认模型。微软表示,Adobe、Figma Weave、Minute(Sports Illustrated)、Wirestock 和 Manus 均与 MAI Image 有合作。
MAI Image 2.7 旨在编辑和分离图像中的单个元素,而不改动其余部分。|图片:Microsoft
与 OpenAI 的 API 兼容的语音模型
MAI Realtime 1 是 Microsoft 的首个语音到语音模型。它旨在应对打断、等到说话人把话说完,并且即便有背景噪声也能识别请求。对于更复杂的任务,它会回退到按 Token 计费的更强大推理引擎。基础费率按音频分钟计费。Microsoft 表示,在 Big Bench Audio、τ-Voice 以及针对 Dynamics 365 的测试中,它的表现优于 GPT Realtime 2.1。
由于它与 OpenAI Realtime API 兼容,现有的语音应用可以轻松切换过来。Agoda 和 Reservations.ai 正在测试该模型。Microsoft 还计划将其引入 Dynamics 365 Contact Center 和 Dragon Copilot,护士可在后者中用其练习与患者的对话。
根据 Microsoft 自身的测评,MAI Realtime 1 在 Big Bench Audio 上以相对较低的成本取得最高分。|图片:Microsoft
原文较长,译文为节选;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 is replacing outside AI models in its own products with cheaper in-house alternatives)
GPT-Image-2 prompted by THE DECODER
Microsoft's AI division unveiled seven specialized models for code, IT security, image editing, speech, and workflow decisions at its Ignite conference. All are aimed at cutting operating costs for always-on AI agents.
The models target OpenAI's and Anthropic's systems as cheaper alternatives. The coding model MAI Code 2 nearly matches GPT-6.1 Sol and Claude Opus 5.5 on internal benchmarks at a fraction of the cost. MAI Cyber 2 Flash edges past GPT-6 Astra on the CyberGym-E2E vulnerability test.
Beyond the image model MAI Image 2.7, the speech model MAI Realtime 1, and the Google Gemma-based routing model Decision 1.1, Microsoft will let enterprise customers fine-tune selected model weights with their own data in the Azure cloud.
Microsoft AI (MAI) is targeting always-on AI agents with new models for code, security, images, speech, and routing. The main yardstick is the lineup of its own partner, OpenAI.
At its Ignite conference, MAI unveiled seven new models, releasing them in a single batch rather than spread across the year, just as it did at Build in June. Microsoft is positioning the new lineup primarily on price-to-performance. The company cites rapidly rising token spending among customers that want to run AI agents around the clock. Its code models target everyday enterprise coding tasks and aren't yet optimized for harder work such as foundational research or long-horizon agentic workflows, Microsoft says.
MAI Code 2 approaches GPT-6.1 Sol and Opus 5.5 at a promotional price
According to Microsoft, MAI Code 2 averages 98 percent of GPT-6.1 Sol's performance across four benchmarks at 40 percent of the cost. Against Claude Opus 5.5, it reaches 96 percent of the performance at 5 percent of the cost. Microsoft tested on DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas, and FrontierCode v1.1. Three of the tests ran on Microsoft's internal MSBench platform using the GitHub Copilot CLI without sub-agents. Microsoft is calculating a 50 percent promotional discount for its own model.
In blind comparisons of frontend work such as UIs, games, data visualizations, and SVG graphics, human evaluators preferred MAI Code 2's output 64 percent of the time over GPT-6 Sol, Microsoft says.
MAI Code 2 trails GPT-6.1 Sol and Opus 5.5 slightly on success rate but costs only a fraction per task, calculated at the promotional price. | Image: Microsoft
The successor to MAI-Code-1.1-Flash, MAI Code 2 Flash, is designed for code explanations, targeted edits, and routine fixes. It's said to match GPT-6 Luna and Claude Haiku 5.5 at 65 and 20 percent of their respective costs.
Microsoft derived MAI Work Flash from the Flash model. It runs in Microsoft 365 and Fabric IQ. On internal Excel and PowerPoint benchmarks, it matches GPT-6 Luna, Microsoft says. In Word, it performs similarly to GPT-6.1 Sol while consuming about 42 percent fewer input tokens. For generating DAX queries in Power BI, it outperforms GPT-5.6 Sol, GPT-6 Astra, and Opus 5.5.
In GitHub Copilot, MAI Code 2 costs $0.50 per million input tokens and $2.50 per million output tokens at the promotional price through May 17, 2027, then $1 and $5. MAI Code 2 Flash costs $0.05 and $0.25 until then, rising to $0.10 and $0.50.
Security model with a cleaned-up benchmark
MAI Cyber 2 Flash solves 92.3 percent of the tasks in the CyberGym-E2E vulnerability benchmark, Microsoft says. GPT-6 Astra reaches 91.5 percent but needs twice the reasoning budget. Per task, the MAI model costs less than 2 percent of GPT-6 Astra, with per-token costs at one hundredth. On WebExploitBench, it scores 40.4 percent, ahead of GPT-5.6 Cyber at 38.4 percent.
MAI Cyber 2 Flash solves slightly more tasks than GPT-6 Astra on CyberGym-E2E while using fewer tokens at a fraction of the cost. | Image: Microsoft
During evaluation, Microsoft found that 10 percent of the roughly 900 CyberGym tasks draw on the same open-source projects and vulnerability reports as parts of its own training data. The reference solutions never entered training, the company says. Microsoft removed those tasks for all models anyway and scored only the rest.
A comparison with GPT-6 Cyber is missing so far. The model is available only to selected customers. Agentic benchmarks with tools still need extra safety measures, Microsoft says. The company plans to publish those results in the coming weeks.
The context window spans one million tokens, four times the size of MAI Cyber 1.1 Flash. Output tokens cost 86 percent less than the predecessor's. MAI Cyber models already handle nearly 40 percent of scans on Microsoft's security platform MDASH, the press release states.
Microsoft lists MAI Cyber 2 Flash at $0.10 per million input tokens and $0.50 per million output tokens, compared with $10 and $50 for GPT-6 Astra. On CyberGym-E2E, that works out to $0.45 per task instead of $34.38.
Microsoft's decision model now runs on Google's Gemma
Microsoft-Decision-1.1 handles routing, classification, prioritization, and workflow control. The model picks from developer-approved actions and returns calibrated confidence scores. When uncertainty is high, apps can escalate instead of acting on a wrong decision. If information is incomplete, the model is designed to abstain.
Its foundation is Google's mixture-of-experts model Gemma 4 26B A4B. The predecessor was built on Qwen3.5-9B. Microsoft trained the new model with supervised fine-tuning and reinforcement learning on hundreds of synthetic and licensed datasets and runs it on a proprietary inference engine.
On the Decision Index v2.1 from Artificial Analysis, Decision 1.1 scores 68 points. That's six more than its predecessor and above Jev 1.14, Gemini Decision 1, and GPT-6 Luna (Decisions). Microsoft also tested it on 102 internal benchmarks with more than 385,000 questions, where it achieves the highest average accuracy among eight decision models.
The model now accepts images, including screenshots. In Copilot, it decides whether a request goes to Copilot or Cowork. It performs as accurately as a large language model but 2.2 times faster and 4.4 times cheaper, Microsoft says. In Cowork, its Auto function controls how much reasoning effort a task receives.
Microsoft charges $0.042 per million input tokens, or $0.05 in the EU data zone. Output tokens are free.
On the Decision Index from Artificial Analysis, Decision 1.1 gains six points over its predecessor and responds more than twice as fast. | Image: Microsoft
Another specialized model appears in the Book of News for Ignite. Microsoft-Retriever-1 gathers relevant information from enterprise documents for agents in Foundry IQ. Instead of running a single search, the model plans its queries, evaluates the results, and issues follow-ups until it has enough material for a well-grounded answer. Microsoft says that's faster and more efficient than previous methods. Retriever-1 launches as a preview. The Model Router in Foundry now also selects automatically from an expanded pool of models.
Image editing just ahead of OpenAI
MAI Image 2.7 leads the Artificial Analysis image-editing leaderboard with 1,215 Elo points from 5,478 votes. It's four points ahead of GPT Image 2.6 Sunburst (max) at one-fifth the cost. For text-to-image generation, the model ranks fourth on that leaderboard. MAI Image 2.7 costs about $35 per 1,000 images. Flash costs about $10.
According to Artificial Analysis, MAI Image 2.7 and 2.7 Flash match GPT Image 2.6 Sunburst's editing quality at much lower prices. | Image: Microsoft
The model can change the color, material, or position of individual objects without altering the rest of the image. It also generates transparent backgrounds. The smaller Flash variant costs less than a third and powers Canvas, a new image editor in MAI Labs. MAI Image has been the default model in Bing Image Creator since June. Adobe, Figma Weave, Minute (Sports Illustrated), Wirestock, and Manus work with MAI Image, Microsoft says.
MAI Image 2.7 is designed to edit and isolate individual elements of an image without altering the rest. | Image: Microsoft
Speech model compatible with OpenAI's API
MAI Realtime 1 is Microsoft's first speech-to-speech model. It's designed to handle interruptions, wait until speakers finish, and recognize requests even with background noise. For harder tasks, it falls back to a more powerful reasoning engine billed per token. The base rate is charged per audio minute. On Big Bench Audio, τ-Voice, and tests with Dynamics 365, it outperforms GPT Realtime 2.1, Microsoft says.
Because it's compatible with the OpenAI Realtime API, existing voice apps can switch over easily. Agoda and Reservations.ai are testing the model. Microsoft also plans to bring it to the Dynamics 365 Contact Center and Dragon Copilot, where nurses would use it to practice patient conversations.
MAI Realtime 1 scores highest on Big Bench Audio at comparatively low cost, according to Microsoft's own measurement. | Image: Microsoft
Dragon Copilot Transcribe, based on MAI Transcribe 2, now handles all doctor-patient transcriptions in Dragon Copilot. That amounts to 10 million conversations per month, Microsoft says. In multilingual workflows, the error rate drops by half. For its own large models, Microsoft now uses the Maia 200 chip, the second generation of its AI accelerator, said to deliver 1.7 times more tokens per watt.
Microsoft charges one US cent per audio minute. Reasoning costs an extra $1 per million input tokens and $5 per million output tokens.
Licensed weights available only through the Microsoft account team
With MAI Licensed Weights, companies can fine-tune MAI models with their own data and run them in their own Azure environment. Data and customizations stay with the customer. Microsoft says it will use them for training its own models only with explicit consent.
The company positions the offering as an alternative to open weights, whose provenance and behavior are hard for enterprises to verify. One example is US carrier AT&T, where a customized model for 5G network operations analysis reportedly outperformed GPT-6.1 Sol on standard tasks at roughly one-fiftieth of the cost. A launch date and pricing are still pending. Interested companies should contact their Microsoft account team. The offering is set to arrive "shortly" through Microsoft's Frontier Company program.
MAI Code 2 and Code 2 Flash are available in GitHub Copilot, with Code 2 Flash also in a limited preview in Foundry. OpenRouter, Vercel, Hermes, Open Code, and Kilo are expected to follow. MAI Image 2.7 and Flash are in public preview in Foundry and on fal, Genspark, OpenRouter, and Vercel. Decision 1.1 is generally available in Foundry, on OpenRouter, and through Vercel's AI Gateway. MAI Realtime 1 can be tried in MAI Labs and is available as a Foundry preview, free until the public preview launches. MAI Cyber 2 Flash is exclusive to Microsoft's security products. In MDASH, it will automatically replace the existing model in the MAI profile for eligible customers starting mid-December.
In April, Microsoft and OpenAI restructured their partnership. OpenAI can now sell its products through any cloud provider, but Azure remains the primary cloud partner. Microsoft holds a non-exclusive license to OpenAI's models and products through 2032.
AI News Without the Hype – Curated by Humans
原文为英文,正文由自动抓取并保留原文。
原文发布于 The Decoder(Microsoft is replacing outside AI models in its own products with cheaper in-house alternatives),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。