Microsoft presented a new stage of its MAI family at Ignite on November 17, 2026, with models for coding, image generation and editing, voice conversations and code security. The announcements include MAI Code 2 and Code 2 Flash, MAI Image 2.7 and its Flash variant, MAI Realtime 1 and MAI Cyber 2 Flash. Access differs between products: availability in Copilot, public preview for images, limited preview for voice, and a security integration scheduled for December.
Alongside these releases, the company announced MAI Licensed Weights, through which organizations will be able to adapt Microsoft models in their own Azure environments. The conference announcements therefore combine specialist models with several forms of integration into applications and agents. Each component has its own access channel, and preview conditions are published separately from the availability of the coding models.
Coding models with distinct tasks and subscription access
MAI Code 2 targets deep debugging, implementations that change multiple files, refactoring and repository-wide reasoning. Microsoft also includes building interfaces, dashboards and visualizations among its intended uses. Code 2 Flash targets code explanations, focused edits, routine fixes and rapid iterations, for workflows with a high volume of requests.

GitHub confirms Code 2 access for Copilot Pro, Pro+, Max, Business and Enterprise. The Flash variant also includes Free and Student plans. Eligible annual subscribers are covered as well. The models appear in the picker across experiences such as Visual Studio Code, Visual Studio, Copilot CLI, the cloud agent, the Copilot app, JetBrains and Xcode. Business and Enterprise administrators control access through model policies; automatic enablement depends on the global option set by the administrator.
According to Microsoft's evaluations, Code 2 achieves 98% of GPT-6.1 Sol's performance at approximately 40% of the cost, averaged across four coding benchmarks. The comparison includes DeepSWE 1.1, TerminalBench 2.1, SWE-Atlas and FrontierCode 1.1. Models were evaluated at maximum reasoning effort, and the Code 2 cost used in the comparison includes its 50% promotional discount. In Copilot, the two models are billed according to usage.
Foundry availability is narrower in the product announcement: Code 2 Flash is in limited-access preview. Microsoft announces future distribution of the coding models through OpenRouter, Vercel, Hermes, OpenCode and Kilo. These channels are described as subsequent steps, while GitHub Copilot access is confirmed at launch.
Image editing in public preview
MAI Image 2.7 and Image 2.7 Flash are available in public preview through Microsoft Foundry. The company publishes indicative figures of approximately $35 per thousand images for the main model and $10 for Flash. The capabilities described include changing an object's color, material or position, preserving consistency across successive edits, and generating transparent backgrounds for visual assets.
Canvas, the new interactive workspace in MAI Labs, uses the Flash model for work starting from text or sketches and for modifying a selected area. The demonstration page allows up to five reference images to be attached. The displayed MAI Labs conditions specify personal, non-commercial use, with sign-in and a minimum age of 18. Demonstration access through Labs and professional integration through Foundry are separate channels.

Voice conversations with reasoning priced separately
MAI Realtime 1 is a speech-to-speech model for conversations involving listening, spoken responses and actions in business systems. Microsoft describes handling interruptions, brief acknowledgments and requests spoken against background noise. Tone and behavior can be adjusted through instructions. Announced compatibility with the OpenAI Realtime API allows developers to evaluate migrating existing voice applications to this model.
The announced base rate is $0.01 per audio minute. Reasoning is billed separately at $1 per million input tokens and $5 per million output tokens. In Foundry, the model is in limited-access preview; according to Microsoft, these preview customers will not be billed until public preview begins. The company also describes integration into Dynamics 365 Contact Center for customer service applications.
The security model is entering Microsoft products
MAI Cyber 2 Flash has a one-million-token context window, four times the size of MAI Cyber 1.1 Flash's. Distribution is announced exclusively through Microsoft security products, including Codename MDASH. For eligible customers, the new model will automatically replace the existing model in the MAI profile starting in mid-December.
Microsoft reports a 92.3% success rate on CyberGym-E2E, compared with 91.5% for GPT-6 Astra, using half the reasoning budget. The company excluded 92 of the 905 tasks because their projects and vulnerability reports overlapped with part of the training data; the reported results cover the remaining 813 tasks. The per-task cost comparison comes from the offline evaluation and includes input, cached input and output tokens.
Access channels described at launch
| Component | Stage or channel |
|---|---|
| Code 2 and Code 2 Flash | Available in GitHub Copilot; Flash in limited Foundry preview |
| Image 2.7 and Flash | Public preview in Foundry |
| Realtime 1 | Limited-access preview in Foundry |
| Cyber 2 Flash | Microsoft products; MDASH replacement from mid-December |
| Licensed Weights | Announced for future Frontier Company engagements |
Licensed weights for the organization's Azure environment
Through MAI Licensed Weights, businesses will be able to adapt models using proprietary knowledge and control access, deployment and upgrade timing in their own Azure environments. Microsoft states that customers retain their data and adaptations, and that using them to train the company's models requires express agreement. The program is announced as coming soon through Frontier Company engagements, with access discussed through the organization's Microsoft account team.
The customization examples presented by the company concern defined tasks. Woven by Toyota used its MISRA fix history to tune Code 2 Flash and reported merge-ready fixes in 44% of cases where its existing models required human intervention. For maintenance report classification, the Alaska Airlines evaluation improved from 81.2% to 91.1% after tuning. These are results from customer-specific evaluations, with qualified employees retaining responsibility for final decisions in the Alaska Airlines case.