Microsoft announces seven artificial intelligence models including coding, image, and voice systems at Ignite 2026
November 20, 2026 · 3 min read
Microsoft announced seven artificial intelligence models on November 17 at its annual Ignite 2026 event. The models span coding assistance, image generation and editing, voice conversation, decision-making, document retrieval, and security vulnerability detection.
GitHub Copilot now offers two coding models. MAI Code 2 handles modifications across multiple files and repository-wide debugging. MAI Code 2 Flash, a lightweight version, suits brief tasks like code explanation and partial rewrites. Both are available across Copilot Pro, Pro Plus, Max, Business, and Enterprise plans, with Flash also available on Free and Student tiers. Business and Enterprise administrators can set usage policies for each model.
Microsoft benchmarked MAI Code 2 against GPT-6.1 Sol and Claude Opus 5.5 across four coding evaluation frameworks. At promotional pricing of half the standard rate, MAI Code 2 costs an average of 0.309 dollars per task, compared with 0.761 dollars for GPT-6.1 Sol and 6.306 dollars for Claude Opus 5.5. MAI Code 2 Flash, benchmarked against GPT-6 Luna and Claude Haiku 5.5, costs 0.052 dollars per task at promotional pricing, versus 0.080 dollars and 0.319 dollars respectively.
For image work, MAI Image 2.7 now generates images from text in addition to its existing editing capabilities. A lightweight version, MAI Image 2.7 Flash, serves high-volume image tasks. Both create images with transparent backgrounds. Pricing through Microsoft Foundry is approximately 35 dollars per 1,000 images for Image 2.7 and approximately 10 dollars for Flash. Microsoft also released Canvas, an image creation tool on MAI Labs that builds images from text instructions or sketches and allows editing of specific elements.
MAI Realtime 1 is Microsoft's first voice conversation model. It listens to human speech and responds vocally, supporting natural conversation features like brief acknowledgments and interruptions. Developers can integrate it with business systems to create voice agents that process spoken requests. The model is compatible with OpenAI Realtime API. Basic voice interaction costs 0.01 dollars per minute, with advanced reasoning at 1 dollar per million input tokens and 5 dollars per million output tokens. Microsoft plans to integrate MAI Realtime 1 into Dynamics 365 Contact Center for customer service and Dragon Copilot for medical training.
Microsoft-Decision-1.1, now generally available, returns probabilities for developer-specified choices, helping applications route processing between lightweight and large models. The model now accepts images and screenshots as input. Pricing through Microsoft Foundry is 0.042 dollars per million input tokens with free output tokens.
Microsoft Retriever 1, in preview, helps agents search internal documents by understanding user questions and constructing search queries using document-specific terminology. Testing showed that configurations using Retriever 1 completed 80 percent of queries without calling external models and reduced median search time from 4.46 seconds to 0.76 seconds, though the share of required facts found decreased from 69 percent to 64 percent.
MAI Cyber 2 Flash, a security model, detects vulnerabilities and generates fixes. It processes 100 million tokens at once, four times the previous generation, with output token costs 86 percent lower. Microsoft will deploy it in Codename MDASH, an autonomous code security system that multiple AI models use to scan code and assist security teams.
GitHub and Microsoft tested MDASH on six open source projects: AutoGPT, Caddy, Express, FastMCP, ImageMagick, and Netty. The scan found 9,791 vulnerability candidates. After filtering for duplicates and validity, Microsoft researchers verified 99 confirmed vulnerabilities and reported them with fixes to project maintainers. By November 15, maintainers had fixed 55 of these, with 14 fixes included in released versions. GitHub plans to expand vulnerability scanning to all 216 projects supported by its Secure Open Source Fund and future fund recipients.
Microsoft Licensed Weights allows enterprises to adjust Microsoft models using proprietary data and run them in their own Azure environments. Enterprises control who accesses the models and when to update versions. Microsoft may use enterprise data and adjustment results for its own model training only with explicit consent. Alaska Airlines adjusted MAI Code 2 Flash to classify aircraft maintenance reports, improving accuracy from 81.2 percent to 91.1 percent.