跳到正文
Hacker News (official RSS)(microsoft.ai)·· 5 天前精選品質評分86

Microsoft 推出 MAI Code 2 與 Code 2 Flash,主打日常程式工作

MAI Code 2 and Code 2 Flash

重點導讀

企業寫程式時,常要在處理複雜任務、速度和成本之間取捨。

  • Microsoft 推出 MAI Code 2,主打除錯、多檔案修改和整理整個程式專案。
  • MAI Code 2 Flash 則適合大量的程式說明、指定修改和例行修正。Microsoft 表示,在四項日常企業程式基準測試中,MAI Code 2 的表現約為 GPT 6.1 Sol 的 98%,成本是 40%。
  • 這些比較會隨工作內容、設定和價格假設而變動。兩款模型已在 GitHub Copilot 提供,Code 2 Flash 也在 Microsoft Foundry 開放限量預覽。

AI 伴隨問答

自備 Key 深度分析

針對這篇報導有疑問?可直接點選快捷問題或向 AI 追問背景脈絡與技術細節:

推薦理由

文中以日常企業程式工作為範圍,比較兩款模型的效能與成本,也列出不適用的高難度工作,可用來判斷它們適合哪些開發任務。

正文 · 原文

Nov 17, 2026 Models

ASCII-art curly braces over an abstract background

: two new coding models built for the everyday enterprise coding tasks, delivering leading performance at a fraction of the cost of competing models.

Performance and speed for everyday coding

MAI Code 2 is at the frontier of price-performance for common enterprise coding tasks and excels in taking code to production, deep debugging, multi-file implementation, repo wide reasoning, and UI. It does all of this while also delivering changes code maintainers would merge, at a fraction of frontier cost.

Meanwhile, MAI Code 2 Flash is designed for high-volume work such as code explanations, targeted edits, routine fixes, and rapid iterations, while bringing quality at a much lower cost.

Production quality at less than half the price

To deliver the best price-performance, we’ve built these models to prioritize the coding work our customers do every day. They’re explicitly not yet tuned for harder problems, like fundamental science research and very long horizon agentic work.

Averaged over four widely used benchmarks that target the everyday enterprise coding tasks we see in GitHub Copilot (DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1), MAI Code 2 gives you 98% of the performance of GPT 6.1 Sol at 40% of the cost, and 96% of Claude Opus 5.5’s performance at just 5% of the cost.1

MAI Code 2 is also tuned for producing great dashboards and visualizations, and excels at visual coding. In blind frontend app evaluations, reviewers preferred its results over GPT-6 Sol 67% of the time.2

In addition, MAI Code 2 Flash offers comparable performance to GPT-6 Luna and Claude Haiku 5.5 at 65% and 20% of the cost, across three coding benchmarks measuring day-to-day, high-volume routine coding tasks.3

Beyond Coding

MAI Code 2 is also becoming the workhorse behind Microsoft products. From Code 2 Flash we further optimized a single model, MAI Work Flash, to serve multiple Microsoft products across M365 and Fabric IQ.

Excel: Our model is on par for accuracy and token efficiency with GPT 6 Luna across key internal Excel customer-scenario benchmarks.

PowerPoint: On key production benchmark scenarios that measure frequent PowerPoint tasks, our model performs on par with GPT 6 Luna in both quality and token efficiency.

Word: Our model performs similar to GPT 6.1 sol (while being significantly smaller) on 5 key internal benchmarks using ~42% fewer input tokens per call than GPT6.1 Sol in production.

Teams: Our model beats GPT 6 Luna at tracking decisions, filling knowledge gaps, and adapting custom instructions using up to 28% fewer tokens on instruction following and tool calling.

FabricIQ: In offline evaluation runs our MAI model outperforms GPT-5.6-sol, GPT-6-astra, and Opus-5.5 at DAX query generation, the core capability behind “Chat with your Data” for Power BI semantic models.

Get started

MAI Code 2 and Code 2 Flash are available in GitHub Copilot, and Code 2 Flash is in limited-access preview in Microsoft Foundry. Soon, we’ll be releasing it on Open Router, Vercel, Hermes, Open Code, and Kilo.

Learn more about customizing MAI Code 2 on our Custom Models page.

About these results
Results reflect evaluations on specific coding tasks and configurations. Performance and cost comparisons vary by workload, model configuration, reasoning effort, and pricing assumptions.

  1. MAI Code 2 performance and price: All models were run at max reasoning effort. DeepSWE 1.1, TerminalBench 2.1 and SWE Atlas were run on MSBench (Microsoft’s internal eval platform) with the GitHub Copilot CLI without subagents. FrontierCode 1.1 scores and costs are from Artificial Analysis; MAI-Code-2 cost is shown at its 50% promotional price.
  2. Visual coding: Blind human side-by-side evaluations of MAI Code 2 and GPT-6 Sol across user interfaces, games, data visualizations, 3D scenes, and SVG graphics. The preference covers both aesthetics and functionality, and is annotated by a total of 258 annotators.
  3. MAI Code 2 performance and price: All models were run at max reasoning effort. DeepSWE 1.1, TerminalBench 2.1 and SWE Atlas were run on MSBench with the GitHub Copilot CLI without subagents. MAI-Code-2-Flash cost is shown at its 50% promotional price.

Related Stories

MAI Cyber 2 Flash: Leading security. Lower cost

Nov 17, 2026

MAI Licensed Weights: Frontier performance you control

Nov 17, 2026

Seven new world class models at scalable price-performance

Nov 17, 2026

Hill-climbing MAI models for GitHub Copilot and Excel

Jul 23, 2026

讀者討論

理性交流 · 即時討論