Microsoft Unveils Five New Models and Announces Progress on Two Others at Ignite 2026
At Ignite 2026, Microsoft unveiled MAI Code 2, MAI Code 2 Flash, MAI Cyber 2 Flash, MAI Realtime 1, and Microsoft-Decision-1.1, while announcing new progress on MAI Image 2.7 and MAI Transcribe 2. MAI Code 2 achieved an aggregate score equivalent to about 98% of GPT-6.1 Sol’s performance across four coding benchmarks, at approximately 40% of its testing cost. Microsoft also upgraded Copilot’s multi-agent collaboration capabilities, Agent 365, and Microsoft IQ. Since its launch in May this year, Agent 365 has registered more than 50 million agents.
Why read this
AI score 77
The article provides a focused overview of the benchmarks, costs, and availability of seven models, facilitating a side-by-side understanding of how Microsoft’s in-house models work in tandem with its enterprise agent platform.
Topics
Body · AI translation
AI-focused media outlet Zhidx reported on November 18 that in the early hours today, Microsoft unveiled five new models and announced new developments for two others at Ignite 2026. Its coding model, MAI Code 2, achieved an aggregate score across four coding benchmarks equivalent to about 98% of GPT-6.1 Sol’s performance, at approximately 40% of its evaluation cost.
The five new models and upgraded versions released this time are MAI Code 2 and MAI Code 2 Flash for coding, MAI Cyber 2 Flash for cybersecurity, the real-time conversational speech model MAI Realtime 1, and the decision model Microsoft-Decision-1.1.
Microsoft also announced new developments for two previously released models. MAI Image 2.7, an image model released on October 27 this year, rose from second to first place on Artificial Analysis’s image-editing leaderboard, and Microsoft also introduced MAI Image 2.7 Flash. Dragon Copilot Transcribe, a medical speech-transcription feature built on MAI Transcribe 2, which was released this September, has been deployed in Microsoft’s healthcare AI product Dragon Copilot. It covers approximately 10 million patient visits per month and has reduced multilingual transcription error rates by about 50%.
Apart from MAI Code 2, several of the other models also perform close to or even better than mainstream competing products. Microsoft said its cybersecurity model MAI Cyber 2 Flash performed on par with GPT-6 Astra in vulnerability-detection tests, while costing less than 2% as much in compute per task. The image model MAI Image 2.7 topped Artificial Analysis’s image-editing leaderboard, scoring four points higher than GPT Image 2.6 Sunburst Max while costing about one-fifth as much.
Microsoft Chairman and CEO Satya Nadella wrote on X that, through the series of technology and product updates released at Ignite, Microsoft hopes to translate technological advances into real-world outcomes that benefit individuals, businesses, and society as a whole.
▲Satya Nadella’s post (Image source: X)
Users can have Autopilot continuously follow up on projects, schedule meetings, and send task reminders, while also working with AI in Workspaces to revise documents and complete projects. Enterprises can use Agent 365 to centrally manage agents from platforms such as OpenAI and Anthropic, review tasks completed by agents, and estimate their business value. Microsoft disclosed that since Agent 365 launched in May this year, more than 50 million agents have been registered on the platform in total.
Microsoft also placed particular emphasis on upgrading its enterprise knowledge-management platform, Microsoft IQ. The platform integrates capabilities including Work IQ, Fabric IQ, Foundry IQ, and Web IQ, and provides more than 350 third-party connectors to help different agents access enterprise knowledge, business data, and operational tools. Microsoft also introduced Business Plugins, enabling enterprises to package their business rules, domain expertise, and operational capabilities for reuse by different agents.
I. Microsoft unveils seven model releases and updates in rapid succession! Coding performance approaches GPT and Claude, with evaluation costs as low as 5% of competing models
Microsoft unveiled five new models and announced new developments for two others. The two coding models and the decision model are now generally available: the coding models have been integrated into GitHub Copilot, while the decision model powers Auto in Cowork.
MAI Code 2 and MAI Code 2 Flash are primarily designed for software development. The former can handle tasks such as complex code debugging and cross-file modifications. According to Microsoft, across four coding benchmarks, MAI Code 2 achieved an aggregate score equivalent to about 98% of GPT-6.1 Sol’s performance, at approximately 40% of its evaluation cost. Compared with Claude Opus 5.5, it achieved about 96% of its performance at approximately 5% of its evaluation cost.
MAI Code 2 Flash is primarily intended for code explanation, localized modifications, and routine bug fixes. Microsoft said the model performed on par with GPT-6 Luna and Claude Haiku 5.5 across three everyday coding benchmarks, with evaluation costs based on promotional pricing equivalent to approximately 65% and 20% of those models’ respective costs. Both models have been integrated into GitHub Copilot, while MAI Code 2 Flash is also available in limited preview through Microsoft Foundry.
▲MAI Code 2 aggregate performance and evaluation cost (Image source: Microsoft)
In cybersecurity, MAI Cyber 2 Flash can help identify software vulnerabilities, analyze threats, and recommend fixes. In the CyberGym-E2E benchmark, the model achieved a task success rate of 92.3%, slightly above GPT-6 Astra’s 91.5%, while costing less than 2% as much in compute per task. Microsoft plans to begin gradually making the model available to eligible customers through its MDASH security-scanning tool in mid-December this year.
▲Official MAI Cyber 2 Flash technical blog evaluation (Image source: Microsoft)
The MAI Image 2.7 image model was released on October 27, primarily for image generation and localized editing. Since its initial release, the model has risen from second to first place on third-party evaluator Artificial Analysis’s image-editing leaderboard, scoring 1,215 points—four points ahead of GPT Image 2.6 Sunburst Max—while costing about one-fifth as much. Microsoft also launched MAI Image 2.7 Flash for batch image generation and the Canvas workspace for interactive image editing. Both image models are now available in public preview through Microsoft Foundry, and users can also try their image-generation and editing capabilities through MAI Labs.
▲MAI Image 2.7 rises to first place on Artificial Analysis’s image-editing leaderboard (Image source: Microsoft)
On the voice front, MAI Realtime 1 supports real-time voice conversations and can be used for scenarios such as AI customer service and voice assistants. Microsoft says the model outperformed GPT-Realtime 2.1 in tests of speech understanding and task execution. Basic voice service starts at $0.01 per minute, with inference fees charged separately. Users can currently try the model through MAI Labs, while developers can apply for limited preview access through Microsoft Foundry.
Another speech-transcription model, MAI Transcribe 2, was released in September this year. Microsoft has now disclosed that Dragon Copilot Transcribe, a medical speech-transcription capability based on the model, is being used in its Dragon Copilot healthcare AI product to handle all doctor-patient conversation transcription. It covers approximately 10 million patient visits per month and has reduced multilingual transcription error rates by about 50%.
The Microsoft-Decision-1.1 decision model is now generally available. It is primarily responsible for information classification, model selection, and task scheduling, and has gained image-input capabilities. The model scored 68 on the Artificial Analysis Decision Index leaderboard, six points higher than its predecessor and ahead of Jev 1.14, Google’s Gemini Decision 1, and OpenAI’s GPT-6 Luna decision edition. Microsoft says it processes decision-making tasks 50 times faster than GPT-6.1 Sol at about 1/79 of the cost, and that more than 4,000 customers are already using it.
Microsoft also announced MAI Licensed Weights, an enterprise model-weight licensing service that will provide eligible businesses with the weights of its internally developed models. Customers will be able to customize the models using internal data and deploy and run them in their own Azure environments. Businesses can independently control model access, while retaining control over their data and model customizations. The service will be rolled out gradually through Microsoft’s Frontier Company partnership program.
II. Copilot upgrades multi-agent collaboration, enabling AI to proactively find work and follow tasks over the long term
Microsoft’s updates this time primarily add multi-agent collaboration and proactive task recommendations to Cowork, while expanding access to the autonomous Autopilot agent to Frontier customers. Copilot’s unified Home interface and its built-in Office in Copilot file-editing capabilities are also now generally available.
Copilot Home, now generally available, is Copilot’s unified work interface, combining Chat for question answering and Cowork for complex tasks in a single entry point. Users can initiate tasks through one input box, with Copilot’s Auto feature automatically selecting the model and reasoning intensity. If a task is relatively complex, the system will suggest handing it over to Cowork while retaining the preceding conversational context, eliminating the need to enter the instructions again.
▲Copilot Home’s unified work interface (Image source: Microsoft)
Copilot Cowork has also upgraded its Auto feature, using Microsoft’s Microsoft-Decision-1.1 model to select the appropriate model and reasoning intensity for each step of a task. It has also added the ability to run multiple sub-agents in parallel, splitting complex tasks among several sub-agents for simultaneous processing.
For example, when a user requests a competitive analysis report, Cowork can have one sub-agent use a Claude model to research competitors and another use a GPT model to conduct a financial comparison, then consolidate the results into a single report. Over the past week, Microsoft has added several models to Cowork, including GPT-6.2 Sol and Claude Fable 5.2. Auto enhancements such as multi-sub-agent orchestration have begun rolling out to Frontier users.
▲Cowork multi-sub-agent collaboration (Image source: Microsoft)
Cowork has also added proactive task recommendations. Through Work IQ, it can use a user’s emails, meetings, and Teams chat history to identify work that can be delegated to AI, such as preparing materials for a client meeting the next day or following up on important emails that have not yet received a response. Once the user confirms, Cowork can begin carrying out the task. This feature is also now available to Frontier users.
Copilot Autopilot previously entered private preview and is now expanding to Frontier customers. Unlike Cowork, which primarily handles tasks assigned by users, Autopilot can run autonomously over extended periods, continuously monitoring Teams channels, following up on emails, and handling recurring tasks.
For example, users can assign a supplier review project to Autopilot, which will develop a plan, schedule meetings, prepare materials, and contact relevant personnel to collect progress updates. Administrators can restrict Autopilot’s access, review its activity logs, and require human approval before important actions.
Workspaces, the previously announced human-AI collaboration space, is also now available to Frontier customers. Employees and agents can share files and project materials in the same space—for example, an agent can draft a document, employees can suggest revisions, and the agent can then continue refining it.
Office in Copilot is now generally available, integrating the editing capabilities of Word, Excel, and PowerPoint directly into the Copilot app. Through conversation, users can create and modify documents, spreadsheets, and presentations, as well as invite colleagues to co-edit them, without switching to separate Office applications.
In addition, the previously introduced Copilot Code is scheduled to become generally available in the coming weeks, enabling users to develop applications, interactive dashboards, and other tools using natural language. Copilot Managed Runtime, which hosts these applications, is now generally available. Edge for Business has also made Cowork’s ability to perform multistep tasks across websites generally available. It can fill out forms and handle workflows on websites that users are authorized to access, while sensitive actions still require user confirmation.
Regarding enterprise adoption, Microsoft disclosed that more than 5,000 people at Eli Lilly use Copilot Cowork each month, including researchers. Some researchers delegate multistep research and analysis tasks to Cowork, using AI to help identify areas worthy of further study before proceeding with experimental work.
III. More than 50 million agents registered in Agent 365 as Microsoft begins helping enterprises manage their AI investments
Microsoft has expanded Agent 365, its enterprise agent management platform, with enhanced third-party management and security controls, while also adding business value analytics tools.
According to Microsoft, enterprises have registered more than 50 million agents in Agent 365 since its official launch in May this year. As the number of registered agents grows, enterprises need not only to understand each agent’s identity, operational status, and access permissions, but are also beginning to focus on their usage costs and actual effectiveness.
For cross-platform management, Agent 365 has further expanded its support for platforms including OpenAI, Anthropic, Google, Salesforce, and Databricks. Supported agents connected to the platform can automatically receive their own Microsoft Entra Agent ID, enabling enterprises to view their origin, owner, and the tools they invoke, as well as implement identity and access management. On some supported platforms, administrators can also review agent activity and task execution logs. These capabilities are currently in preview and will be rolled out in phases.
For local agent management, Microsoft has expanded its device- and network-level discovery and control capabilities. Tools including Microsoft Defender and Microsoft Entra Global Secure Access can now identify more than 60 types of local AI agents, including Claude Code and GitHub Copilot CLI. Defender’s device-level discovery capability is now generally available, while Global Secure Access’s network-level discovery capability remains in preview. Administrators can discover agents that employees are using independently and, through the relevant management and security tools, restrict their access to enterprise data in supported scenarios.
The newly launched Agent 365 Business Analytics can help enterprises analyze AI usage costs and estimate business value. The tool brings agent usage, expenses, and task completion data into a single interface, providing metrics such as task success rates, user satisfaction, assisted work hours, and estimated business value.
▲Agent 365 Business Analytics dashboard (Image source: Microsoft)
Using Copilot Cowork as an example, the system can combine the amount of human time a task would typically require with the proportion of the work handled by AI to estimate the number of work hours assisted by AI and the corresponding value. However, this is not a direct measurement of human time saved, nor can it be equated with financial gains already realized by the enterprise. Business Analytics is now beginning to roll out to Frontier users.
Previously, Agent 365 Cost Management already offered AI cost management capabilities. At this year’s conference, Microsoft further upgraded its spending analytics and budget controls. Enterprises can view spending by employee, user group, agent, and model, set spending limits and cost alerts, and restrict teams from using high-cost models. For example, administrators can define the range of models available to different teams, reserving high-cost models for tasks requiring complex reasoning and preventing costs from spiraling out of control as AI usage scales up.
IV. Microsoft upgrades Microsoft IQ: More than 350 connectors enable agents to operate business systems directly
According to Microsoft, nearly 60,000 customers are now using Microsoft IQ, its enterprise knowledge platform. Launched at the Build conference in June this year, the platform now integrates capabilities including Work IQ, Fabric IQ, Foundry IQ, and Web IQ, and provides more than 350 third-party connectors. This upgrade focuses on enabling enterprises to centrally manage dispersed knowledge, business rules, and operational tools so they can be reused by different AI agents.
▲Work IQ knowledge and context architecture (Image source: Microsoft)
The newly added Business Plugins allow enterprises to package business knowledge, operational rules, and tools into plugins for reuse by different departments and agents. For example, in Microsoft’s supply chain disruption scenario, an agent can look up affected orders, check supplier contracts, and then identify the employee responsible for handling the issue. Once an enterprise configures the relevant knowledge and operational rules, multiple agents can use them. The feature is now available to Frontier program users.
This time, Work IQ primarily showcased its ability to move from retrieving workplace information to taking action. On November 3, Microsoft opened a public preview of more than 60 native SharePoint actions. Agents can not only query project materials but also directly perform tasks such as creating project tracking lists and processing approvals. Microsoft’s internal testing showed that after these tools were integrated, Work IQ’s task pass rate rose from 70.1% to 84.6%, while token usage fell by 16%.
The Listen feature can automatically initiate tasks when a message is received or a business event occurs, without waiting for the user to issue an instruction. Listen is currently in public preview.
Microsoft has also extended Work IQ to third-party AI assistants. According to Microsoft, ChatGPT Business and Enterprise users can now use a plugin to access workplace information in Outlook emails, Teams chats, and SharePoint files after receiving authorization. A plugin for Claude will also be rolled out gradually.
In terms of enterprise business system partnerships, Microsoft and Salesforce announced the Copilotforce partnership, which will bring Salesforce customer data and business capabilities into Copilot. The newly launched Salesforce in Copilot plugin includes more than 38 prebuilt sales skills. For example, salespeople can combine customer records in Salesforce with communications from Outlook and Teams to query sales opportunities and prepare materials for customer meetings. The plugin is currently in public preview.
Conclusion: Microsoft connects a complete suite of enterprise agent tools as its in-house models compete for business
At this Ignite conference, Microsoft further expanded its enterprise AI strategy. On the one hand, Microsoft upgraded Copilot into a unified work hub that integrates agent execution, enterprise knowledge, and management tools. On the other hand, Microsoft prominently showcased the performance and cost advantages of its in-house AI models, beginning to compete directly with partners such as OpenAI and Anthropic in the model business. Microsoft aims to gain greater control over enterprise AI workflows while continuing to offer third-party models.
For Microsoft, its existing Office user base, Azure cloud services, and enterprise data connectivity provide the foundation for Copilot to enter real-world business operations. However, whether enterprises can continue scaling their use of agents will still depend on the reliability of task execution, the effectiveness of cross-platform collaboration, and actual returns on investment. Microsoft’s introduction of cost and business value analysis tools also reflects how competition in enterprise AI is moving beyond adding features toward controlling costs and validating real-world results.
Source: Microsoft
AI Trends