Microsoft Ignite 2026 Recap: MAI Coding and Security Models, Copilot Agents, Azure AI Foundry and What Platform Teams Should Do

Microsoft Ignite 2026 Recap: MAI Coding and Security Models, Copilot Agents, Azure AI Foundry and What Platform Teams Should Do

Microsoft Ignite 2026 Recap: MAI Coding and Security Models, Copilot Agents, Azure AI Foundry and What Platform Teams Should Do

Microsoft spent the opening day of Microsoft Ignite 2026 telling enterprises that its own models can now do most of what OpenAI and Anthropic models do for a fraction of the price. MAI Code 2 went live in GitHub Copilot on November 17, with Microsoft claiming 96% of Claude Opus 5.5’s coding performance at 5% of the price. That claim lands in a week where an independent count found that only 76 of 215 newly announced features were generally available and free to use on day one. For platform teams, the gap between those two facts is the whole story: the model economics are shifting fast, while most of the governance and agent plumbing is still in preview. This recap sorts the verified announcements from days one to three by status, scores our own pre-event predictions, and turns the noise into an adoption matrix and a 30/60/90-day plan.

What this covers: the MAI model lineup, Copilot and Agent 365 changes, Microsoft Foundry, security and identity for agents, infrastructure and data, a prediction scorecard, risks, and a concrete checklist.

Status note: Ignite 2026 runs November 17–20 in San Francisco. This recap reflects announcements verified through November 19. The final day may add news, and preview features listed here can change before general availability.

Context and Background

Ignite is Microsoft’s annual enterprise conference, and since 2024 it has been the venue where the company resets its agent and Copilot roadmap for the coming year. In 2026 it runs November 17–20 at Moscone Center in San Francisco, with the opening keynote by Satya Nadella on day one. Microsoft publishes a “Book of News” on that first morning as the canonical list of announcements, and that document is the primary source for most of this recap.

Our Microsoft Ignite 2026 preview for platform teams made a deliberate choice: it labelled every expected item as analysis, not announcement, and warned that “if these items don’t land, nothing in your plan should depend on them.” That framing matters now, because the useful question after a keynote is not “what was said” but “what can I actually turn on, in which region, under which licence.”

Three structural shifts frame this year’s event. First, Microsoft AI (the division led by Mustafa Suleyman) has moved from shipping a few in-house models to shipping a portfolio. Earlier in 2026 it released speech models, which we covered in our explainer on MAI-Voice-2.1 and MAI-Transcribe-2 pricing, and earlier coding models such as MAI-Code-1-Flash, whose upcoming deprecation GitHub announced in an August 11, 2026 changelog. At Ignite 2026 the division announced seven new models in a single post.

Second, the product formerly marketed as Azure AI Foundry now appears in Microsoft’s materials simply as Microsoft Foundry. Many teams still search for “Azure AI Foundry agents”, so this article uses both names; they refer to the same platform.

Third, the volume of announcements now outpaces anyone’s ability to evaluate them. Office Watch counted 215 separately named features across 59 Book of News items and found that 76 (35%) were generally available at no extra cost on day one, 8 were available at extra cost, and 109 were preview or Frontier-program only. Fourteen more were announced without a date. That ratio is the single most useful number for any team planning a 2027 roadmap around Ignite: roughly six in ten announcements are not yet production commitments.

Reuters framed the headline as Microsoft taking on Anthropic and OpenAI with coding and cybersecurity models. Suleyman described the effort as “a hill-climbing machine that we have built, which co-optimizes performance and price,” and said, “This isn’t just one model. It’s not a flash in the pan.” The strategic read is clear: Microsoft wants model choice inside its own stack, with its own models as the cost floor.

Microsoft Ignite 2026 Announcements: What Actually Shipped

The verified Microsoft Ignite 2026 announcements cluster into five layers: in-house MAI models led by MAI Code 2 (GA in GitHub Copilot), a unified Copilot experience with the Autopilot agent rolling out to Frontier customers, Microsoft Foundry evaluation and routing upgrades, agent identity and governance through Entra and Agent 365, and Cobalt 200 VMs reaching general availability.

Microsoft Ignite 2026 stack map showing Copilot, Foundry, MAI models, routing, AKS agent infrastructure, silicon and the Agent 365 governance plane

Figure 1: How the Microsoft Ignite 2026 announcements stack, from silicon to the Copilot experience, with Agent 365, Entra and Perception acting as a governance plane across layers.

Figure 1 places each announcement on a single stack. At the bottom sit Cobalt 200 Arm CPUs and Maia accelerators. Above them, AKS gains an Agent Sandbox extension and a centralized policy layer called Agent Grid, both in preview. The model layer combines MAI models, partner models and two routers. Foundry wraps models with agents, evaluation and Foundry IQ knowledge. The Copilot experience sits on top. Agent 365, Entra Agent ID and Perception cut across every layer, which is the part most teams will spend 2027 wiring up.

MAI models: Microsoft’s own cost floor

The Microsoft AI post “Seven new world-class models at scalable price-performance,” dated November 17, 2026, is the primary source for the model lineup. The verified facts break down as follows.

MAI Code 2 is the general-purpose coding model. Microsoft says that, averaged across DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1, it 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. In a blind frontend evaluation, developers preferred it over GPT-6 Sol 64% of the time. Microsoft also says plainly that it is not yet tuned for fundamental science research or very long-horizon agentic work. The GitHub changelog of November 17 confirms availability in Copilot Pro, Pro+, Max, Business and Enterprise, across VS Code, Visual Studio, JetBrains, Xcode, Eclipse, Copilot CLI, the GitHub Copilot app, the cloud agent, github.com and GitHub Mobile.

MAI Code 2 Flash targets high-volume routine work: explanations, targeted edits and quick fixes. Microsoft says it is comparable to GPT-6 Luna at 65% of its price and to Claude Haiku 5.5 at 20% of its price, measured across DeepSWE 1.1, Terminal Bench 2.1 and SWE-Atlas. It is available on every Copilot plan including Free and Student, and in a limited preview in Microsoft Foundry.

MAI Cyber 2 Flash is the security model. Microsoft says it beats GPT-6 Astra’s quality at under 2% of the cost per task on the CyberGym-E2E benchmark, and that its output tokens cost 86% less than MAI Cyber 1.1 Flash. The Book of News lists it as preview only inside Codename MDASH, Microsoft’s agentic code-security system. It is not a general-purpose API you can call from Foundry today.

MAI Image 2.7 and MAI Image 2.7 Flash handle image generation and editing. Microsoft says MAI Image 2.7 ranks first on Artificial Analysis for image editing and fourth for text-to-image. The Flash variant costs about $10 per 1,000 images. Both are listed as preview in Foundry.

MAI Realtime 1 is Microsoft’s first live speech-to-speech model, starting at $0.01 per audio minute and available to try through MAI Labs via an interest form. Dragon Copilot Transcribe, built on MAI Transcribe 2, is live in Dragon Copilot; Microsoft says it covers about 10 million clinical encounters per month.

Microsoft-Decision-1.1 is a routing model that sends requests to larger models only when needed. It is GA in Foundry, OpenRouter and Vercel, adds image input, and is built on Gemma 4 26B A4B according to the Book of News. Microsoft reports a score of 68 on the Artificial Analysis Decision Index v2.1, six points above the first version, at the same price and more than twice the speed. Microsoft says more than 4,000 customers use it.

Two further items matter for regulated customers. MAI Licensed Weights lets selected customers customize a frontier MAI model with their own data and run it in their own Azure environment. Microsoft also cited AT&T building a custom MAI model for 5G network analytics. Neither has public pricing.

A critical caveat applies to every comparison above: these are vendor-reported, aggregated benchmark ratios against vendor-chosen competitors. Microsoft did not publish per-benchmark scores in the sources we reviewed, and it compared against specific model variants. Treat “96% at 5% of the price” as a hypothesis to test on your own repositories, not as a procurement fact.

Copilot becomes an agent platform, not an assistant

The Copilot announcements are less about chat and more about long-running agents with their own identities. The official Microsoft blog recap of Nadella’s keynote describes Copilot combining Chat, Cowork, Code and Autopilot into one experience, which it calls “the new OS for work.”

Autopilot is the headline agent. The Book of News defines it as a long-running personal productivity agent with its own identity, memory and tool access, able to run recurring work and be @mentioned in Teams. Its actions are tied to its own agent ID. It is rolling out through the Frontier program, Microsoft’s early-access channel, so it is not GA.

Copilot Managed Runtime is GA, with an optional Fabric data backend. Copilot SDK 2.0 is GA, with a native Rust runtime, support in six programming languages and multi-tenant enterprise security; the Book of News says the SDK “uses the same harness technology deployed in production.” Copilot Code, which lets users build dashboards, trackers and apps in natural language inside a sandbox, is in Frontier now with general availability expected in December. Cowork is GA as part of the new Home experience, and Microsoft says nearly 80% of the Fortune 500 use it.

On model choice, the Copilot blog says employees can let Auto routing pick a model and effort level per task, or manually choose OpenAI or Anthropic models subject to organizational policy. Some reports before and during the event spoke of 20 or more frontier models available in the Copilot harness. We could not confirm a specific model count in the Book of News, the Copilot blog or the keynote recap, so treat any number as unverified.

Microsoft Foundry: evaluation and routing grow up

Foundry’s Ignite 2026 story is about operating agents, not building them. The Book of News lists multi-turn evaluation, insights on production signals and the agent optimizer as generally available. The optimizer “proposes and tests candidate changes” against evaluations, which is the closed loop many teams have been building by hand. Human annotation with golden datasets, Managed Training and Programmed Training for custom supervised fine-tuning and reinforcement-learning loops are in preview.

The Foundry model catalog now lists more than 11,000 models, with the Book of News naming OpenAI, Claude, Muse, Mistral and the open-source community among the sources. The model router, GA since Ignite 2025, will draw on “the top 50 models in Foundry,” work with AI Gateway and support more Messages and Responses APIs. Notebookcheck notes that no availability date was given for the expanded pool. Foundry IQ adds knowledge graphs and a Microsoft-Retriever-1 model in preview, plus a GA SharePoint indexed knowledge source.

Infrastructure for agents also moved. You can now deploy Foundry open-weight models and agents to your own AKS clusters (preview). An Agent Sandbox extension for AKS offers VM-isolated code execution (preview), and Agent Grid promises centralized policy across agent, model, tool and service interactions (preview). If you have been comparing isolation options, our guide to sandboxing AI agent code execution with Firecracker, gVisor and WebAssembly explains what “VM-isolated” buys you over container isolation.

Announcement status table

Announcement Layer Status (per primary source) Platform-team action
MAI Code 2 Model GA in GitHub Copilot Run a repo-level bake-off now
MAI Code 2 Flash Model GA in GitHub Copilot; limited preview in Foundry Default for routine edits; Foundry API in sandbox only
MAI Cyber 2 Flash Model Preview, only inside Codename MDASH Evaluate via MDASH if eligible
MAI Image 2.7 / 2.7 Flash Model Preview in Foundry Non-production creative pilots
MAI Realtime 1 Model Try via MAI Labs, interest form Watch list
Microsoft-Decision-1.1 Routing GA in Foundry, OpenRouter, Vercel Test as a cost router
MAI Licensed Weights Model access Announced, selected customers Engage account team if sovereign needs
Autopilot Copilot agent Rolling out to Frontier Identity and audit design first
Copilot Managed Runtime Copilot GA Assess against custom hosting
Copilot SDK 2.0 Developer GA Pin versions, plan upgrades
Copilot Code Copilot Frontier; GA expected December Sandbox and data policy review
Agent 365 Cost Management Governance Spend visibility and limits in preview Enable alerts in pilot tenant
Entra Agent ID for local agents Identity Preview Pilot with Copilot CLI and Claude Code
Microsoft MCP Server for Enterprise Identity GA Review permissions model
MCP Firewall in Global Secure Access Network security Preview Design MCP allow-lists
Microsoft Perception Security ops GA Map to SOC runbooks
Codename MDASH AppSec Preview, eligible orgs Pilot on one repo group
Foundry multi-turn eval, agent optimizer Foundry GA Wire into CI
Agent Sandbox for AKS, Agent Grid Infrastructure Preview Non-production evaluation
Azure Cobalt 200 VMs Compute GA Benchmark Arm migration candidates
Windows Server 2028 OS Preview Lab only
Azure HorizonDB Data GA Evaluate for PostgreSQL workloads

The table compresses roughly 40 of the 215 features into the ones most likely to change a platform team’s roadmap. Everything marked preview or Frontier should be read as “subject to change in name, scope, price and region.”

Deeper Analysis: Model Economics, Agent Identity and Infrastructure

The individual announcements matter less than the system they form. Three threads deserve a closer look: how MAI models change the cost structure of coding agents, how agent identity is being bolted onto the Microsoft security stack, and what the infrastructure and data updates mean for teams running workloads outside Copilot.

Why routing, not a single model, is the real MAI story

Microsoft did not ship one model to replace GPT or Claude. It shipped a ladder: a Flash tier for volume, a full tier for harder work, and a router that decides which to call. Figure 2 shows how that ladder works inside a coding-agent loop.

Microsoft MAI models routing flow showing Decision 1.1 sending coding requests to MAI Code 2 Flash, MAI Code 2 or a frontier partner model with an evaluation gate

Figure 2: A routing loop for coding agents. A router sends routine edits to MAI Code 2 Flash, multi-file work to MAI Code 2, and long-horizon tasks to a partner frontier model, with evaluations and cost telemetry feeding back into routing.

In this loop, the router is the policy point. Decision 1.1 or the Foundry model router classifies each request. Routine edits go to the cheapest adequate model, multi-file refactors go to MAI Code 2, and the work Microsoft itself says MAI Code 2 is not tuned for, such as long-horizon agentic tasks, goes to a frontier partner model. A multi-turn evaluation gate decides whether output is accepted, and cost telemetry from Agent 365 Cost Management closes the loop.

The economics follow from the ratios Microsoft published. Consider an illustrative team whose coding-agent traffic is 70% routine edits and 30% complex work, currently all sent to a premium frontier model. If Microsoft’s price ratios held for that workload, moving routine traffic to a model priced at 20% of the incumbent’s small tier, and complex traffic to a model priced at 5% of the incumbent’s flagship, would cut spend dramatically. However, this arithmetic is illustrative only. It assumes identical token counts per task, identical retry rates and identical acceptance rates, and none of those are guaranteed.

The hidden variable is the retry rate. A model at 96% of the quality may fail some tasks that the incumbent passes, and each failure costs another call, a human review or a revert. If a cheaper model needs two attempts on one task in ten, its effective cost rises by roughly 10% before you count engineer time. That is still cheap at a 5% price ratio, but the reviewer minutes are not. Our fact-check of whether AI coding agents really make developers faster explains why measured productivity often diverges from benchmark gains.

Gartner’s August 2026 forecast, cited by Notebookcheck, adds context: inference cost per agentic workflow is predicted to rise more than fivefold through 2028 even as per-token prices fall, because agents make many more calls. Gartner recommends inference tiering, which is exactly the pattern in Figure 2. Microsoft is effectively productizing Gartner’s advice, with its own models occupying the cheap tiers.

There is also a billing nuance. The GitHub changelog states that both MAI Code models “are billed under usage-based billing,” but it does not list a per-token rate or a premium-request multiplier on the page itself; it points to GitHub’s models-and-pricing documentation. Before you forecast savings, pull the current rate from that documentation rather than from secondary reporting.

Agent identity: the security stack catches up with local agents

The most consequential security change at Ignite 2026 is that Microsoft now treats agents running on developer laptops as first-class identities. Entra Agent ID for discovered local agents is in preview, and it starts with GitHub Copilot CLI and Claude Code. The goal, per the Book of News, is to separate agent activity from human activity in logs and policy.

Copilot frontier models agent identity sequence showing a local agent obtaining an Entra Agent ID token and calling an MCP server through Global Secure Access MCP Firewall with Agent 365 logging

Figure 3: The intended control path for a local coding agent. Intune discovers the agent, Entra issues an agent-scoped token, Global Secure Access applies MCP Firewall policy, and Agent 365 and Defender record the activity.

Figure 3 strings together five separate announcements into one control path. Intune’s local AI agents governance (preview) discovers agents on endpoints and can mark devices with unapproved agents as noncompliant. Entra issues a token tied to the agent rather than the user. Global Secure Access can selectively acquire agentic traffic (GA) and apply an MCP Firewall (preview) that controls access to MCP servers, tools, prompts and transports. Agent 365 and Defender record activity, and Defender runtime protection for local agents is GA on Windows for supported agents, requiring Defender for Endpoint P2.

The pattern mirrors what the identity community has been designing in the open. Our deep dive on AI agent identity and delegated authorization with OAuth token exchange covers the on-behalf-of flows that make an agent-scoped token meaningful. The Microsoft version adds endpoint discovery and network enforcement, which open standards do not cover.

Two caveats temper the excitement. Most of the path is preview, so the end-to-end chain is not production-supported today. And the chain assumes Microsoft owns the endpoint, the identity provider and the network edge. Teams using Okta, a non-Microsoft secure web gateway or Linux developer machines will get only fragments.

On the security-operations side, Microsoft Perception (formerly Project Perception) is GA. It is a multi-agent system with red, blue and green team agents and playbooks. Microsoft says its Triage Agent covers about 80% of alerts in a typical Defender queue and now handles AWS and Salesforce alerts. Codename MDASH, in preview for eligible organizations, uses more than 100 specialized agents in a four-stage pipeline to map the attack surface, scan code, validate findings and remove duplicates. Microsoft says MAI Cyber models account for almost 40% of MDASH scans.

Purview also extends to agents. Copilot data loss prevention (preview) can block Copilot from referencing sensitive files even when unlabeled, and endpoint agent data security (preview) controls what agents on Windows can read and share. Agent 365 governance adds plugin and skills governance (GA), MCP server and tool governance (preview) and local and shadow agent governance (preview), with an action center entering private preview in December.

Infrastructure, silicon and data

Azure Cobalt 200 VMs are GA. Microsoft says they deliver up to 50% better CPU performance than Cobalt 100, offer consistent-speed and boost modes, and add Mpsv4 for large in-memory databases and Lpsv5 with up to 23 TB of storage. The Book of News lists 11 regions with existing deployments, including East US, West US 2, UK South and Sweden Central, and says deployment will reach 24 regions by year end. Our earlier comparison of Cobalt 200, Graviton and Axion Arm silicon remains the place to start an Arm migration assessment.

The accelerator story is less tidy. The Microsoft AI post lists Maia 200 as Microsoft’s second-generation accelerator, in production running its largest models, with 1.7x more tokens per watt. The official keynote recap on the Microsoft blog announces Maia 300 with 40% more tokens per watt than leading GPU-based systems. The Book of News sections we reviewed do not mention Maia at all. The most consistent reading is that Maia 200 is in production and Maia 300 is the next generation, but no customer-facing Maia instance or price was announced, so this does not change any planning decision yet.

Elsewhere in infrastructure, Azure SRE Agent gained customer-managed keys, Entra Agent ID and Agent 365 integration, cross-agent communication and multicloud investigations (preview). Flex nodes for AKS let clusters extend to on-premises servers (preview). Windows Server 2028 entered preview, with hybrid AI features in limited early access and support for AMD MI350P and NVIDIA GB300 accelerators. Confidential computing gained three previews, including VR72 multi-GPU confidential computing for models larger than one GPU.

In data, Azure HorizonDB is GA with Entra ID, private endpoints, managed identities and customer-managed keys, while its AI model management and AI pipelines are preview. Fabric HorizonDB, a serverless PostgreSQL with pause and resume, is preview. Database Hub in Fabric is GA, and Fabric IQ integrations across Copilot and Fabric apps are GA, with ChatGPT and Claude connectors in preview. Microsoft IQ itself is GA with no new SKU. For Azure SQL, next-generation General Purpose zone redundancy and flexible memory are GA, and a free local-development container is in preview.

Scorecard: which of our preview predictions held

Our preview post listed six expected items and labelled all of them as analysis. Here is how they fared against the Book of News and Microsoft blogs.

Preview prediction Outcome Evidence
Copilot harness roadmap to GA or harness switching Partly held Copilot Managed Runtime and Copilot SDK 2.0 GA; SDK uses the production harness; no harness-switching announcement found
Single agent inventory across Agent 365 and Foundry Not confirmed Book of News mentions an Agent Registry inside Dynamic Agent Risk, preview; no unified Foundry inventory stated
AI Gateway tier progress and pricing Partly held Model router to work with AI Gateway; Azure AI Gateway support for Agent 365 MCP governance in preview; no gateway pricing found
Evaluation and agent optimizer GA Held Multi-turn evaluation, production insights and agent optimizer listed GA
Copilot SDK package and namespace cut-over dates Not confirmed SDK 2.0 GA; no package or namespace names in the Book of News
Foundry Local on Azure Local GA Not confirmed Foundry Local appears as a Windows Server 2028 runtime and on Windows 365 Cloud PCs; no Azure Local GA statement found

Two of six predictions held or mostly held, two partly held, and two could not be confirmed. The preview also missed the biggest story: it said nothing about MAI models or chips. That miss is instructive. Platform roadmaps built only from product-team signals underweight what Microsoft AI, a separate organization, ships.

Adoption decision matrix

Use the matrix below to decide what to act on this quarter. “Adopt” means production use with guardrails, “Pilot” means a bounded trial on non-critical workloads, and “Watch” means no engineering time yet.

Capability Status Lock-in risk Value if you are Microsoft-centric Value if you are multi-cloud Recommendation
MAI Code 2 in GitHub Copilot GA Low, switchable in model picker High High Pilot now, adopt after bake-off
MAI Code 2 Flash in Foundry API Limited preview Medium Medium Medium Watch
Decision 1.1 router GA, also on OpenRouter and Vercel Low High Medium Pilot
Foundry evals and agent optimizer GA Medium, eval assets portable if exported High Medium Adopt for Foundry agents
Autopilot Frontier High, tied to M365 identity Medium Low Watch, design identity now
Entra Agent ID plus MCP Firewall Preview High High Low Pilot in one business unit
Microsoft Perception GA High, Defender-centric High Medium, AWS and Okta coverage added Pilot in SOC
Codename MDASH with MAI Cyber 2 Flash Preview Medium High Medium Pilot if eligible
Cobalt 200 VMs GA Low to medium, Arm portability High Medium Adopt for suitable stateless tiers
Agent Sandbox for AKS, Agent Grid Preview Medium Medium Medium Watch

The matrix makes one pattern obvious. The GA items with low lock-in are the developer-facing ones, MAI Code models in Copilot and the Decision router. The governance items with the highest strategic value are also the least mature and the most tightly coupled to Microsoft identity and endpoint products.

Trade-offs, Gotchas, and What Goes Wrong

The first risk is preview churn. With roughly six in ten Ignite 2026 features in preview, Frontier or undated, any architecture that depends on them can break when names, APIs or prices change. Microsoft’s own naming moved during this cycle: the platform once called Azure AI Foundry is now Microsoft Foundry, and Project Perception became Microsoft Perception. Expect SDK namespaces and portal paths to follow.

The second risk is inconsistent primary sources. We found at least three discrepancies across Microsoft’s own pages. MAI Code 2 Flash is “GA on GitHub Copilot and in Foundry” in the corporate keynote recap, but limited or private preview in Foundry according to the Microsoft AI post and the Book of News. The Microsoft AI post cites Maia 200 while the keynote recap cites Maia 300. The model post also refers to both “GPT-6.1 Sol” and “GPT-6 Sol” as comparison points. When pages disagree, the Book of News and product changelogs should win, because they carry status labels.

Preview-to-production gate for Microsoft Ignite 2026 announcements routing each feature by GA, preview, Frontier or undated status

Figure 4: A preview-to-production gate. Each Ignite announcement is routed by its status label, then by whether licensing and billing are known, then by whether an exit path exists if the feature changes.

Figure 4 encodes the rule we recommend: nothing reaches production without a GA label, a known licence and billing model, and an exit path. Public previews get a sandbox pilot with no production data. Private previews, Frontier features and undated announcements stay on a watch list.

The third risk is licensing stacking. Several capabilities carry prerequisites that only appear in footnotes. Defender runtime protection for local agents requires Defender for Endpoint P2. Agent 365 Business Analytics needs a Microsoft 365 Copilot licence for Microsoft insights and an Agent 365 licence for non-Microsoft services. Viva Insights removed the Agent 365 licence requirement but keeps a 50-Copilot-licence minimum, and full governance requires Agent 365 or ME7. Today Feed’s recurring updates consume Copilot credits. Budget for the full stack, not the headline SKU.

The fourth risk is lock-in through governance. Model choice is getting more open: Copilot exposes OpenAI and Anthropic models, and Decision 1.1 runs on OpenRouter and Vercel. Governance, however, is getting more closed, because agent identity, MCP firewalling and endpoint discovery all assume Entra, Intune and Global Secure Access. That asymmetry is deliberate. Teams that want portability should keep evaluation datasets, traces and policy definitions in formats they control, such as OpenTelemetry traces, which our guide to OpenTelemetry GenAI semantic conventions covers in depth.

The fifth risk is benchmark overreach. Every model comparison at Ignite 2026 came from Microsoft and used aggregated ratios against chosen competitors. Our survey of AI agent benchmarks such as SWE-bench, GAIA and tau-bench explains why aggregate scores hide task-level variance. A model that averages 96% may be far weaker on your language, framework or monorepo layout.

Finally, cost controls lag capability. Agent 365 Cost Management offers spend visibility and limits in preview, with forecasting only in Frontier. Long-running agents like Autopilot will run before mature spend controls are GA, which inverts the safe order of operations.

Practical Recommendations

Treat Ignite 2026 as two separate programs. The first is a model-cost program you can start this week, because MAI Code 2 is GA in a surface you already govern. The second is an agent-governance program that you design now but deploy as features reach GA in 2027. Keep them separate so that preview churn in governance never blocks cheap wins in coding.

Assign one owner to reconcile the Book of News against your tenant. Every claimed feature needs a status, region and licence check before it enters a plan. Re-run that reconciliation after the event closes on November 20, and again in January, when several previews (the Agent 365 action center and Perception-based AI Readiness Initiatives among them) are scheduled to move.

30/60/90-day checklist

First 30 days

  • [ ] Export the Book of News into a tracker and tag each item GA, preview, Frontier or announced-only.
  • [ ] Confirm your Copilot model policy: decide whether MAI Code 2 and Code 2 Flash stay enabled by default, since new models auto-enable unless admins disable them.
  • [ ] Run a two-week bake-off of MAI Code 2 against your current default model on real tickets; measure acceptance rate, retries and review minutes.
  • [ ] Pull current MAI Code rates from GitHub’s models-and-pricing documentation and set budget alerts.
  • [ ] Inventory local agents such as Copilot CLI and Claude Code on developer endpoints.

Days 31–60

  • [ ] Pilot Decision 1.1 or the Foundry model router on one internal agent with a measured baseline.
  • [ ] Wire Foundry multi-turn evaluation into CI for any Foundry agent you already run.
  • [ ] Stand up a sandbox tenant to trial Entra Agent ID for local agents and the MCP Firewall preview.
  • [ ] Benchmark one stateless service on Cobalt 200 in a region where it is deployed.
  • [ ] Map licence prerequisites for Agent 365, Defender for Endpoint P2 and Copilot credits.

Days 61–90

  • [ ] Decide adopt, pilot or watch for each item in the decision matrix, with written exit criteria.
  • [ ] Draft an agent-identity standard covering naming, owners, scopes and audit retention.
  • [ ] Re-check statuses of Copilot Code (GA expected December) and the Agent 365 action center.
  • [ ] Export agent traces to an OpenTelemetry store you control.
  • [ ] Present a 2027 roadmap that funds GA items and parks previews.

Frequently Asked Questions

What were the biggest Microsoft Ignite 2026 announcements?

The headline items were seven new MAI models from Microsoft AI, led by MAI Code 2 and MAI Code 2 Flash, both live in GitHub Copilot on November 17. Microsoft also announced the Autopilot personal agent for Frontier customers, GA for Copilot Managed Runtime and Copilot SDK 2.0, Microsoft Perception security agents at GA, Entra Agent ID for local agents in preview, and general availability of Azure Cobalt 200 VMs. Most governance features remain in preview.

What is MAI Code 2 and how much does it cost?

MAI Code 2 is Microsoft AI’s general-purpose coding model for debugging, multi-file implementation, refactoring and repository-wide reasoning. Microsoft says it reaches 98% of GPT-6.1 Sol’s performance at about 40% of the price, and 96% of Claude Opus 5.5’s at 5% of the price, averaged across four coding benchmarks. In GitHub Copilot it is billed under usage-based billing. The changelog does not list a per-token rate, so check GitHub’s models-and-pricing documentation for current figures.

Is Azure AI Foundry the same as Microsoft Foundry?

Yes. Microsoft now markets the platform as Microsoft Foundry, and Ignite 2026 materials use that name throughout. It covers the model catalog of more than 11,000 models, Foundry Agent Service, Foundry IQ knowledge features, evaluation tools and the model router. Documentation and community posts still use “Azure AI Foundry,” so search for both terms. The agent and evaluation capabilities announced at Ignite 2026 apply to the same platform regardless of name.

Which Copilot frontier models can enterprises choose at Ignite 2026?

Microsoft’s Copilot blog says Auto routing picks a model and effort level per task, and employees can also manually choose OpenAI or Anthropic models subject to organizational policy. GitHub Copilot adds MAI Code 2 and MAI Code 2 Flash to its model picker. Microsoft also said GPT 6.2 and Meta’s latest models are coming to Foundry, without dates. We could not verify a specific total model count in Copilot from primary sources.

What is MAI Cyber 2 Flash and can I use it?

MAI Cyber 2 Flash is a security-focused model for vulnerability discovery, validation and remediation. Microsoft says it beats GPT-6 Astra’s quality at under 2% of the cost per task on CyberGym-E2E. According to the Book of News, it is in preview only inside Codename MDASH, Microsoft’s agentic code-security system for eligible organizations. It is not offered as a general Foundry API, and Reuters reported that Microsoft plans to offer the cyber model to a limited group of customers.

When does Microsoft Ignite 2026 end, and will more be announced?

Ignite 2026 runs November 17–20, 2026, in San Francisco. Most announcements arrived with the Book of News on day one, but sessions on the final day may add detail, documentation and dates. Several items also have scheduled follow-ups, including Copilot Code GA expected in December and the Agent 365 action center entering private preview in December. Re-check statuses after the event and again in January.

Further Reading

References

  1. Microsoft. “Microsoft Ignite 2026 Book of News.” November 17, 2026. https://news.microsoft.com/events/ignite/2026/book-of-news/
  2. Microsoft AI. “Seven new world-class models at scalable price-performance.” November 17, 2026. https://microsoft.ai/news/microsoft-ignite-2026/
  3. GitHub Changelog. “MAI Code 2 and MAI Code 2 Flash in GitHub Copilot.” November 17, 2026. https://github.blog/changelog/2026-11-17-mai-code-2-and-mai-code-2-flash-in-github-copilot
  4. The Official Microsoft Blog. “Ignite 2026: Building the frontier ecosystem.” November 17, 2026. https://blogs.microsoft.com/blog/2026/11/17/ignite-2026-building-the-frontier-ecosystem/
  5. Microsoft Copilot Blog. “Copilot at Microsoft Ignite: Expanding what’s possible with intelligence built for work.” November 17, 2026. https://www.microsoft.com/en-us/copilot/blog/2026/11/17/copilot-at-microsoft-ignite-expanding-whats-possible-with-intelligence-built-for-work/
  6. Reuters, via The Korea Times. “Microsoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAI.” November 18, 2026. https://www.koreatimes.co.kr/world/20261118/microsoft-unveils-ai-models-for-coding-cybersecurity-as-it-takes-on-anthropic-openai
  7. Office Watch. “Microsoft Ignite 2026: only 76 of 215 new features are ready today.” https://office-watch.com/2026/microsoft-ignite-2026-ready-today/
  8. Notebookcheck. “Microsoft expands Foundry model router at Ignite 2026 as costs climb.” November 18, 2026. https://www.notebookcheck.net/Microsoft-expands-Foundry-model-router-at-Ignite-2026-as-costs-climb.1453248.0.html

By Riju — about

Comments

No comments yet. Why don’t you start the discussion?

Leave a Reply

Your email address will not be published. Required fields are marked *