MAI (Microsoft AI) is Microsoft’s family of fully in-house models for software development, reasoning, images, speech and cybersecurity. At Ignite 2026 on 17 November 2026, Microsoft AI announced seven new models, including MAI Code 2, MAI Code 2 Flash, MAI-Image-2.7, MAI Realtime 1 and MAI Cyber 2 Flash. Unlike the open Phi models, all MAI models are proprietary and run through Microsoft services: Microsoft Foundry, Azure Speech, GitHub Copilot or Microsoft security products.
This page is based on the Microsoft AI Ignite overview, the Ignite 2026 Book of News, the model overview in Microsoft Learn, the GitHub price list and the Azure Retail Prices API. As an AI consultancy from Rosenheim we classify the models for enterprises in the DACH region by access route and EU processing. This overview reflects the state as of 19 November 2026.
Current models at a glance
| Model | Type | Status | Access route | EU per Microsoft |
|---|---|---|---|---|
| MAI Code 2 | Coding | GA since 17 November 2026 | GitHub Copilot | no EU processing commitment documented |
| MAI Code 2 Flash | Coding | GA in Copilot, limited access preview in Foundry | GitHub Copilot, Microsoft Foundry | Foundry: only Global prices so far |
| MAI-Image-2.7 / -2.7-Flash | Image generation and editing | Public preview | Microsoft Foundry, fal, OpenRouter, Vercel | Global Standard |
| MAI Realtime 1 | Speech-to-speech | Limited access preview | Microsoft Foundry, MAI Labs | no details |
| MAI Cyber 2 Flash | Cybersecurity | Rollout from mid-December 2026 | Codename MDASH | no details |
| MAI-Thinking-1 | Reasoning LLM | Public preview | Microsoft Foundry | Global Standard, all EU regions |
| MAI-Transcribe-2 | Speech-to-text | Public preview | Azure Speech | North Europe, West Europe |
| MAI-Transcribe-2-Streaming | Real-time speech-to-text | Public preview | Azure Speech | Sweden Central |
| MAI-Voice-2.1 / -2.1-Flash | Text-to-speech | Public preview | Azure Speech | France Central, North Europe, Sweden Central, West Europe |
| MAI-Image-2.6 / -2.6-Flash | Image generation | GA | Microsoft Foundry | Global Standard, Sweden Central, West Europe |
We cover the decision model Microsoft-Decision-1.1, also part of the Ignite announcements, on its own page Microsoft-Decision-1.1.
MAI Code 2 and MAI Code 2 Flash
MAI Code 2 targets everyday enterprise development work: production-ready code, debugging, multi-file changes, repo-wide reasoning and frontend work. Per Microsoft the model reaches 98% of GPT-6.1 Sol’s performance at about 40% of the price and 96% of Claude Opus 5.5’s performance at 5% of the price, averaged over DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1. In a blind test with 258 raters, developers preferred its frontend results over GPT-6 Sol in 64% of cases. Per Microsoft all models ran at maximum reasoning effort; Microsoft measured DeepSWE, Terminal Bench and SWE-Atlas on its internal platform MSBench via the Copilot CLI, while FrontierCode comes from Artificial Analysis. Microsoft calculates costs at the promotional price. Per Microsoft the model is not tuned for fundamental science research or very long-horizon agentic work.
MAI Code 2 Flash is meant for high-volume routine tasks: code explanations, targeted edits, simple fixes and fast iterations. Microsoft cites performance comparable to GPT-6 Luna at 65% and to Claude Haiku 5.5 at 20% of the price. The Foundry model catalog lists the model as mai-code-2-flash (version 2026-11-17) with a 1M-token context, up to 128K output tokens, and text and image input. Microsoft has also derived MAI Work Flash from Code 2 Flash, which Microsoft uses in Microsoft 365 and Fabric IQ products.
Per the GitHub changelog and the GitHub pricing list, both models have been generally available in GitHub Copilot since 17 November 2026: MAI Code 2 on Pro, Pro+, Max, Business and Enterprise, MAI Code 2 Flash also on Free and Student. They are available among others in VS Code, Visual Studio, JetBrains, Xcode, Eclipse, the Copilot CLI, the Copilot cloud agent and on github.com. In Business and Enterprise, new models are enabled by default unless administrators have turned off the global default. In Microsoft Foundry only MAI Code 2 Flash is available, as a limited access preview.
MAI-Image-2.7 and MAI-Image-2.7-Flash
MAI-Image-2.7 and MAI-Image-2.7-Flash generate and edit images. Microsoft first showed 2.7 on 27 October 2026 as an image editing model outside Foundry; since Ignite both models are available as a public preview in Microsoft Foundry (model version 2026-11-02), and also via fal, Genspark, OpenRouter and Vercel. Per Microsoft, MAI Image is now the default model family of Bing Image Creator.
Per Microsoft, MAI-Image-2.7 ranks #1 on the Artificial Analysis image editing ranking AA-Image-Editing v2.1 with 1,215 Elo from 5,478 votes (as of 16 November 2026), 4 Elo ahead of GPT Image 2.6 Sunburst (max) at one fifth of the price. For text-to-image Microsoft cites rank 4. The models process a prompt of up to 32,000 tokens and up to five reference images, change the colour, material or position of individual objects in a targeted way and generate images with transparent backgrounds. Per Microsoft Learn, photorealistic edits involving minors are blocked by default.
MAI Realtime 1
MAI Realtime 1 is Microsoft AI’s first speech-to-speech model. It is designed to hold natural conversations with interruptions and quick acknowledgements, hand harder tasks to a stronger reasoning stage, understand requests despite background noise and connect to enterprise systems and tools. The model is compatible with the OpenAI Realtime API, which is meant to ease the migration of existing voice applications. Microsoft states that MAI Realtime 1 beats GPT Realtime 2.1 on quality, speed and price, but publishes no concrete values.
Per Microsoft the price starts at USD 0.01 per audio minute; reasoning costs USD 1 per 1M input tokens and USD 5 per 1M output tokens. Enterprises can register for the limited access preview in Microsoft Foundry and pay nothing until the public preview. A demo runs in MAI Labs, which per its terms of use is intended for personal, non-commercial purposes only. Microsoft plans to use it in Dynamics 365 Contact Center and Dragon Copilot.
MAI Cyber 2 Flash
MAI Cyber 2 Flash finds vulnerabilities in code, proves them and supports the investigation and detection of attacks. Microsoft offers the model exclusively in Microsoft security products, including Codename MDASH, for verified defenders. For eligible customers it replaces the existing model in the MDASH MAI profile from mid-December 2026; the Book of News refers to a preview.
Per Microsoft, MAI Cyber 2 Flash reaches 92.3% on CyberGym-E2E versus 91.5% for GPT-6 Astra, with half the reasoning budget and under 2% of the cost per task; Microsoft excluded 92 of the 905 tasks for all models because of overlap with training data. On WebExploitBench Microsoft cites 40.4% versus 38.4% for GPT-5.6 Cyber. The context grows fourfold over MAI Cyber 1.1 Flash to 1M tokens, and the model produces 86% fewer output tokens. The list price is USD 0.10 per 1M input tokens and USD 0.50 per 1M output tokens.
MAI Licensed Weights
MAI Licensed Weights is meant to let enterprises customize an MAI model with their own data and run it in their own Azure environment, with control over deployment, access and updates. Per Microsoft, data and adaptations remain with the customer; Microsoft uses them for training only with express agreement. As examples Microsoft cites AT&T (customized versions of MAI Code 2 and Code 2 Flash for 5G network analytics), Woven by Toyota (ready-to-merge fixes in 44% of the cases where other models needed rework) and Alaska Airlines (classification accuracy from 81.2% to 91.1%). The offering is announced and comes through Frontier Company engagements; Microsoft has not yet published dates, prices or licence terms. These are not open weights: the models still run in Azure.
MAI-Thinking-1
MAI-Thinking-1 (public preview since 2 June 2026) is the family’s reasoning model: sparse mixture-of-experts with 35B active parameters, a 256k context and up to 64k output tokens, trained on commercially licensed data per Microsoft. Microsoft cites 97.0% on AIME 2025, 94.5% on AIME 2026 and 52.8% on SWE-Bench Pro. In Microsoft Foundry the model runs via an OpenAI-compatible chat completions API (/mai/v1) with function calling and encrypted chain of thought, exclusively as Global Standard. The Azure price list shows USD 2.00 per 1M input tokens, USD 0.20 for cached input and USD 8.00 for output.
MAI-Transcribe-2 and MAI-Transcribe-2-Streaming
MAI-Transcribe-2 (public preview since 3 September 2026) transcribes 60 languages including German with speaker diarization, word timestamps, keyword biasing, automatic language identification and code switching; the style is selectable (verbatim with filler words or clean). Per Microsoft the average word error rate on FLEURS is 5.2%. The model is called via the Azure Speech fast transcription API (enhancedMode.model = "MAI-Transcribe-2") or as input transcription in Voice Live. The price is USD 0.10 per audio hour as a promotional offer until 31 December 2026. Dragon Copilot Transcribe is built on MAI-Transcribe-2; per Microsoft it transcribes all doctor-patient conversations in Dragon Copilot, about 10 million encounters a month.
MAI-Transcribe-2-Streaming is the real-time variant for call centers, voice assistants, live captions and real-time notes. It connects via the Azure Speech SDK or the Realtime API (OpenAI-Realtime-compatible WebSocket protocol). Per Microsoft Learn, Azure serves the model from Sweden Central, Central US, East US 2 and Southeast Asia.
MAI-Voice-2.1 and MAI-Voice-2.1-Flash
MAI-Voice-2.1 and MAI-Voice-2.1-Flash are the family’s text-to-speech models (public preview) with 23 languages, emotion and style control via SSML and instant voice cloning from 5 to 60 seconds of reference audio, which is available only after limited-access approval and with consent. The Flash variant targets low-latency voice agents, IVR and call centers, MAI-Voice-2.1 long-form content such as audiobooks, podcasts and educational content. For German, Microsoft Learn lists the voices de-DE-Klaus, de-DE-Mia, de-DE-Grant and de-DE-Harper.
Pricing
| Model | Input | Cached input | Output | Access and note |
|---|---|---|---|---|
| MAI Code 2 | 0.50 | 0.025 | 2.50 | GitHub Copilot, promotional price until 17 May 2027, then 1.00 / 0.05 / 5.00 |
| MAI Code 2 Flash | 0.05 | 0.005 | 0.25 | GitHub Copilot and Foundry (Global), promotional price until 17 May 2027, then 0.10 / 0.01 / 0.50 |
| MAI-Code-1.1-Flash | 0.20 | 0.02 | 1.20 | GitHub Copilot until 18 December 2026 |
| MAI-Thinking-1 | 2.00 | 0.20 | 8.00 | Microsoft Foundry, Global Standard |
| MAI Cyber 2 Flash | 0.10 | 0.01 | 0.50 | List price per Microsoft AI, only in Microsoft security products |
All prices in USD per 1M tokens. The image and speech models use their own units:
| Model | Price | Source |
|---|---|---|
| MAI-Image-2.7 | about USD 35 per 1,000 images; per 1M tokens: text input 5.00, image input 8.00, image output 34.00 USD | Microsoft AI, Azure price list (Global) |
| MAI-Image-2.7-Flash | about USD 10 per 1,000 images; per 1M tokens: 1.75 / 2.50 / 10.00 USD | Microsoft AI, Azure price list (Global) |
| MAI-Image-2.6 / -2.6-Flash | per 1M tokens: 5.00 / 8.00 / 38.00 USD and 1.75 / 2.50 / 19.00 USD | Azure price list (Global) |
| MAI Realtime 1 | from USD 0.01 per audio minute, reasoning 1 / 5 USD per 1M tokens | Microsoft AI |
| MAI-Transcribe-2 | USD 0.10 per audio hour until 31 December 2026 | Azure Speech pricing page, Microsoft AI |
| MAI-Voice-2.1 | per 1M characters per the Azure Speech pricing page | Azure Speech pricing page |
Prices as of 19 November 2026. The Azure price list also shows Data Zone prices for EU regions for MAI-Image-2.7 (for example USD 40.80 per 1M image output tokens instead of USD 34.00) and an hourly price of USD 0.54 for MAI-Transcribe-2. Because Microsoft Learn does not yet document the Data Zone deployment type for MAI-Image and Microsoft does not explain whether USD 0.54 is the price after the promotion ends, we check such values directly in the Foundry portal in projects.
EU availability and GDPR
For the GDPR assessment, access route and deployment type matter, not the model family:
| Access route | Models | Processing per Microsoft |
|---|---|---|
| Microsoft Foundry, Global Standard | MAI-Thinking-1, MAI-Image-2.5 to 2.7 | Inference possible in any Azure region; stored data stays in the chosen geography |
| Azure Speech, fast transcription | MAI-Transcribe-2, MAI-Transcribe-1.5 | Resource in North Europe or West Europe; per Microsoft Learn, Azure Speech processes in the resource region |
| Azure Speech, Voice Live | MAI-Transcribe-2, MAI-Transcribe-2-Streaming | MAI-Transcribe-2 in EU regions with processing scope Data Zone, streaming in Sweden Central with processing scope Global |
| Azure Speech, text-to-speech | MAI-Voice-2.1, -2.1-Flash | Served from France Central, North Europe, Sweden Central and West Europe, among others, with routing between these regions |
| GitHub Copilot | MAI Code 2, MAI Code 2 Flash | no EU processing commitment documented for the models |
| Microsoft security products, MAI Labs | MAI Cyber 2 Flash, MAI Realtime 1 (demo) | no region details; MAI Labs for personal use only |
For Foundry models Microsoft states: prompts and responses are not used to train the base models and are not available to other customers or model providers. Even with Global deployments, stored data, including the abuse monitoring data store, stays in the customer-designated geography; for deployments in the EEA, per Microsoft only employees located in the EEA review flagged content. New since 17 November 2026: for models made available on or after 1 November 2026, Microsoft intends to stop offering the Modified Abuse Monitoring program from a date not yet announced, with at least 30 days’ notice. Based on the version dates this affects MAI-Image-2.7 and MAI-Code-2-Flash, among others. In addition, per Microsoft preview models may follow different privacy practices.
Our assessment: for personal data, the speech models with a resource in an EU region are currently the most suitable option, for transcription for example MAI-Transcribe-2 in West Europe or North Europe. We use Global Standard models such as MAI-Thinking-1 and MAI-Image for content without personal data, or assess them case by case with a data protection impact assessment and a data processing agreement.
Model lifecycle
- MAI-Code-1.1-Flash is deprecated across all GitHub Copilot experiences on 18 December 2026 per the GitHub changelog; the replacement is the latest MAI Code Flash model, i.e. MAI Code 2 Flash.
- MAI-Code-1-Flash has been retired in GitHub Copilot since 10 September 2026.
- MAI-Image-2.5, -2.5-Flash and -2.5-Pro retire on 24 November 2026 per the Microsoft Foundry retirement schedule (as of 18 November 2026). Successors are MAI-Image-2.6 and MAI-Image-2.7.
- MAI-Image-2e has been retired since 28 August 2026 per the retirement schedule; MAI-Image-2 is no longer listed there.
- MAI-Transcribe-2-Streaming is listed as a preview with a retirement date of 4 December 2026.
- MAI-Transcribe-1 has been retired since 15 September 2026; successors are MAI-Transcribe-1.5 and MAI-Transcribe-2.
- MAI-Voice-2 and -2-Flash are no longer listed in the current Microsoft Learn documentation; successors are MAI-Voice-2.1 and -2.1-Flash.
- MAI Cyber 1.1 Flash is replaced by MAI Cyber 2 Flash for eligible MDASH customers from mid-December 2026.
- Apart from the coding models in GitHub Copilot, most MAI models run as previews without an SLA.
Our recommendation
For coding assistants, MAI Code 2 in GitHub Copilot is worth testing when cost per developer matters: the promotional prices until May 2027 are well below those of the large frontier models. MAI Code 2 Flash suits routine tasks and teams on Copilot Free or Student. If you have pinned MAI-Code-1.1-Flash, switch before 18 December 2026. For code with high confidentiality requirements, we review Copilot policies and data flows first.
For image generation and product images, MAI-Image-2.7 in Microsoft Foundry is worth evaluating, and the Flash variant at about USD 10 per 1,000 images for high volumes. For transcription in the Microsoft stack, MAI-Transcribe-2 with a Speech resource in West Europe or North Europe remains our first choice for evaluations; we plan it for production only once it is generally available. For voice agents we are watching MAI Realtime 1 until Microsoft names languages, regions and a public preview. We use MAI-Thinking-1 for reasoning tasks without personal data.
Via the AI Gateway and CompanyGPT, MAI models from Microsoft Foundry can be connected centrally alongside other models, with budgets and model switching in one place. innFactory GmbH is a Microsoft Solutions Partner; we support selection, evaluation, data protection review and integration.
Related Microsoft models
Open models for self-hosting and edge are covered on the Microsoft Phi page, the decision model for routing and classification on the Microsoft-Decision-1.1 page. We cover the OpenAI models in Microsoft Foundry under OpenAI GPT.
