Microsoft announced seven new first-party AI models at Ignite on November 17, 2026, spanning coding, image generation, cybersecurity and decision-making. The immediate releases include MAI-Code-2 and MAI-Code-2-Flash in GitHub Copilot, plus Microsoft-Decision-1.1 in Microsoft Foundry. Other models remain in preview, including a cybersecurity model available only within Codename MDASH, according to Microsoft’s Ignite Book of News.
The distinction matters for anyone planning an integration. A model available inside Copilot is not automatically available through a Foundry deployment, and a generally available security product does not automatically expose its underlying models.
Decision 1.1 also deserves a closer look because it works differently from a conversational LLM. Its documented API accepts bounded questions and returns numerical decisions, giving developers a way to classify, route and rank inputs without asking for a written response.
Which Models Are Available, and Where?
Microsoft describes the announcement as a seven-model release. The breakdown below separates the individually named models from Microsoft Perception, the model-backed security system, so the product’s availability is not confused with direct model access.
| Model or system | Main task | Announced availability |
|---|---|---|
| MAI-Code-2 | Coding | Generally available in GitHub Copilot |
| MAI-Code-2-Flash | Coding | Generally available in GitHub Copilot; private preview in Foundry |
| MAI-Image-2.7 | Image generation | Preview |
| MAI-Image-2.7-Flash | Image generation | Preview |
| MAI-Cyber-2-Flash | Cybersecurity | Preview, only in Codename MDASH |
| Microsoft-Decision-1.1 | Classification, routing, ranking, grading and binary decisions | Generally available in Foundry |
| Microsoft Perception | Security system using specialized agents and models, including the MAI-Cyber line | Generally available as a system |
The model availability comes from Microsoft’s Book of News and Decision 1.1 documentation. Microsoft separately describes Perception’s general availability in its Ignite overview.
My view is that the access channel is more useful than the model count. For a team building an application, the consequential question is whether it can deploy a model through an API today, use it within an existing product, or only evaluate it through a preview.
Coding Models Ship in Copilot, With Narrower Foundry Access
MAI-Code-2 and MAI-Code-2-Flash are generally available in GitHub Copilot. Microsoft lists Code-2-Flash separately as a private preview in Foundry.
Those are different delivery paths. Copilot availability places the coding models inside Microsoft’s developer assistant. Foundry access concerns deploying a model for an application or workflow. A team using a coding model in Copilot should therefore avoid assuming it can immediately build its own service around the same model.
The distinction is particularly clear for Code-2-Flash: its Copilot status is GA, while its Foundry status is private preview. The announcement does not support treating both coding models as generally available Foundry offerings.
The Flash branding also should not substitute for a workload comparison. Choosing between coding models requires evidence about the tasks that matter to a team, such as producing usable changes, following repository conventions or completing a particular workflow. The availability announcement establishes where these models can be accessed; it is not an independent comparison of coding quality.
Image Models Remain in Preview
MAI-Image-2.7 and MAI-Image-2.7-Flash are preview releases, according to the Book of News. They extend Microsoft’s first-party image-generation lineup, but they do not share the coding models’ GA status in Copilot.
For an enterprise evaluating either model, the next step is to check the preview’s actual access conditions, supported deployment surface and applicable terms. The broad preview designation alone does not answer those questions.
I would keep an image-model evaluation separate from a production migration decision. A useful evaluation can establish whether generated images fit a particular workflow, but that still leaves operational questions about access and service commitments. The announcement’s mixed availability makes that separation especially important.
Perception Is GA; MAI-Cyber-2-Flash Has Restricted Access
MAI-Cyber-2-Flash is available in preview only in Codename MDASH. That is the narrowest explicitly stated model-access path in this release.
Microsoft Perception is generally available, but it is a security system, not a synonym for a standalone MAI-Cyber endpoint. Microsoft says Perception brings together specialized agents to find, defend against and remediate vulnerabilities, powered by models including its MAI-Cyber line. Its Ignite overview describes the product at that system level.
Both availability statements can therefore be true at once: customers can have access to a finished security system while a particular underlying model remains restricted to a preview environment.
For security buyers, the practical evaluation should follow the layer they are purchasing. Assessing Perception means examining the system’s permissions, findings, remediation behavior and fit with existing security operations. Evaluating a cyber model directly means examining its model interface and permitted uses. General availability at one layer does not establish access at the other.
Decision 1.1 Returns Decisions, Not Written Answers
Microsoft-Decision-1.1 is generally available in Foundry for classification, routing, ranking, grading and binary decisions. Its technical documentation describes inputs covering text, JSON and images.
Microsoft says version 1.1 adds multimodal support, improves accuracy and response speed, and builds on a model used by more than 4,000 customers. The company also says it uses the model for routing, data labeling, personalization and content filtering. Those adoption and performance statements are Microsoft-reported; the published sources here provide no independent performance measurements or like-for-like cost comparison.
The most consequential documented detail is the output contract. Decision 1.1 does not generate a free-form answer or written rationale. An application supplies a state and one or more typed questions, and the model returns structured decisions.
The state can contain text or JSON and can reference uploaded images supplied through assets. The three question types determine what comes back:
| Question type | Purpose | Returned result |
|---|---|---|
noul | Test a yes-or-no condition | A probability from 0 through 1 |
choice | Select one option from a fixed set | The selected option and probabilities for every option |
score | Assess an item on an ordered scale | A score and probabilities for the scale’s levels |
This design moves part of the application logic into an explicit decision interface. Instead of asking a chat model to write a department name and then parsing its response, a developer can define the permitted departments as options.
That reduces output-format handling. It leaves the separate task of checking whether the decision is correct.
Use noul for Gates and Filters
A noul question asks whether a condition is true. Microsoft’s examples include detecting whether a tool call touches production data, a document contains prompt injection, or a support response promises a refund.
Frequently Asked Questions
3 questions
1Are All the New Microsoft Models Generally Available?
No. The coding models are GA in GitHub Copilot, and Decision 1.1 is GA in Foundry. The image models remain in preview, while MAI-Cyber-2-Flash is preview-only in Codename MDASH.