Microsoft unveils MAI Code 2 coding model at Ignite


Microsoft says the programming model nearly matches rivals at a fraction of the cost. The company also released image-generation and transcription models and plans limited access to a cybersecurity model.
Microsoft unveiled MAI Code 2, a programming model, at its Ignite conference in San Francisco on 17 November 2026. It leads a group of seven new superintelligence (SI) models that the company says are intended to lower costs for developers and businesses running automated tasks.
The launch is part of Microsoft’s effort to challenge rivals including Anthropic and OpenAI, according to a Reuters report carried by CNA.
Microsoft says MAI Code 2 comes close to rival models on coding tests at a fraction of their cost. It also released image-generation and transcription models and plans to offer a cybersecurity-specific model to a limited group of customers. GitHub has separately added an optional scan for exposed secrets in code changes.
The launch follows an earlier Microsoft announcement of seven models at Build on 2 June 2026, which included MAI-Code-1-Flash alongside models for images, transcription and other tasks.
Microsoft says MAI Code 2 is designed for work including debugging and changes across multiple files. MAI Code 2 Flash targets routine, high-volume work such as explaining code and making focused fixes. Neither model is yet tuned for fundamental science research or very long-running autonomous tasks, the company says.
Microsoft’s comparison averages performance across four tests of everyday enterprise coding: DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1.

For Flash, Microsoft says performance was comparable to GPT-6 Luna and Claude Haiku 5.5 across three routine coding tests, at 65% and 20% of their respective costs.
GitHub says MAI Code 2 is available through Copilot Pro, Pro+, Max, Business and Enterprise. Flash is also available on the Free and Student plans. Both are billed based on usage.
Microsoft AI chief Mustafa Suleyman told Reuters that specializing models for different fields helps Microsoft keep costs low while matching rivals and sometimes exceeding them on tests.
Suleyman co-founded DeepMind and Inflection AI before becoming Microsoft AI’s chief executive in March 2024.
MAI Image 2.7 and MAI Image 2.7 Flash generate and edit images and are in public preview on Microsoft Foundry. Microsoft lists approximate prices of $35 per 1,000 images for the former and $10 per 1,000 images for Flash. It is also launching Canvas, an interactive image workspace powered by Flash.
Microsoft says MAI Image 2.7 ranks first on Artificial Analysis’s image-editing assessment, ahead of GPT Image 2.6 Sunburst (max), at one-fifth of its cost.
Microsoft describes MAI Cyber 2 Flash as a specialist security model built on its security data and expertise. It says the model matched GPT-6 Astra’s quality at less than 2% of the cost per task on the CyberGymE2E test and is being brought into Microsoft security products.
Microsoft says MAI Cyber 2 Flash is designed to find and validate code vulnerabilities and help defenders investigate and detect attacks.
The company says verified defenders can access MAI Cyber 2 Flash through its Codename MDASH security system, with safeguards monitoring and controlling use. Separately, Microsoft plans a limited customer offering for a cybersecurity model intended to help address software vulnerabilities that SI tools are bringing to light.
Suleyman said safeguards were a priority alongside performance, citing OpenAI agents' hacking of the Hugging Face platform. His top priority, he said, was alignment: preventing SI systems from taking harmful actions and ensuring they behave as intended.
GitHub’s new secret-scanning preview adds its detectors to Copilot’s existing security review. It checks proposed code changes, not a repository’s history, and is off by default. Eligible users need no additional plan, but scans consume credits under the Secret Protection AI Credit SKU. The preview works in Copilot’s command-line tool and app sessions, software development kit integrations and code review.
A separate public preview called Adaptive selects a review effort level based on the complexity and risk of a proposed code change. Balanced remains the default, and administrators' maximum settings still apply. Copilot can also use related public repositories for review context in a preview; these features and secret detection are available on Pro, Pro+, Max, Business and Enterprise plans.
Copilot code review for Azure Repos has moved from preview to general availability, with usage billed in GitHub AI credits to the linked Azure subscription. In the Copilot app, Azure DevOps integration is in public preview for code changes and proposed changes known as pull requests; cloud sessions are unavailable for those repositories. Support for GitHub Enterprise Server has also moved to general availability in app version 1.2.0 or later, without cloud sessions for those repositories.
The app’s browser-use preview can interact with pages in a signed-in Chromium-based browser through the Copilot Chrome extension, in an isolated group of tabs. It is off by default. For JetBrains users, GitHub is rolling out previews of an isolated cloud workspace and optional remote access to shared chats. The workspace runs Copilot sessions on GitHub-hosted infrastructure rather than the user’s computer, allowing Copilot to change code and run commands and tests. Browser use through GitHub Copilot in Chrome and on-device voice dictation are also in preview for JetBrains users. Dictation works on supported native plugin packages with Editor Preview enabled; dictated text stays editable until sent.
GitHub has ended Copilot support for JetBrains IDE 2025.2; users need 2025.3 or later to keep receiving updates.
Copilot Fusion, which selects and coordinates models based on a user’s chosen effort level, is available on github.com, Visual Studio and GitHub Mobile beta apps for iOS and Android. GitHub is adding Claude Sonnet 5.5 to Fusion Max, GPT-6.1 Terra to High and MAI-Code-2 to Standard. Support in the production GitHub Mobile apps is coming shortly, GitHub says.
GitHub says Fusion Max resolved 71.2% of tasks on Terminal-Bench 4.1, a test of terminal-based tasks. It says that was 5.4 percentage points ahead of Opus 5.5 and Fable 5.2, at 57% less cost than Opus and 88% less than Fable.
GitHub Copilot Evals is in public preview for Individual, Business and Enterprise plans. It lets users repeat tests of customized skills, agents and tools to compare behavior before and after changes. The extensions and plugins for Visual Studio Code, Copilot CLI and the Copilot app use the local Vally evaluation engine, and users approve each run.
Enterprise Cloud customers with Enterprise Managed Users can stream Copilot agent activity into Microsoft Agent 365, where administrators can review prompts, responses and tool calls across Copilot clients.
Microsoft Entra sign-in for the Copilot command-line tool is in public preview on Windows and macOS. Depending on an organization’s entitlements, it can authenticate a linked managed user or provide access through a Microsoft 365 Copilot license. It requires Copilot CLI version 1.0.113 or later; GitHub says Linux support will follow.
For linked managed users, the single-sign-on route requires an eligible Business or Enterprise license, an enabled command-line policy and an organization using OIDC, a sign-in protocol. It uses GitHub inference governed by the organization’s GitHub Copilot plan. Organizations in Frontier with an eligible Microsoft 365 Copilot license and usage-based billing can instead use Microsoft 365 model processing without a GitHub account; their Microsoft 365 settings control access and limits.

Microsoft says MAI Licensed Weights will let organizations customize and deploy its models in Azure. It also introduced MAI Realtime 1 for enterprise voice conversations, with pricing advertised from $0.01 per audio minute.
4 min