Dev.to AI 🤖 Ai 👁 0 📖 14 min read

Microsoft Ignite 2026: Every Announcement, Sorted (Models, GitHub Copilot, Azure, Windows & More)

Microsoft Ignite 2026 ran November 17–20 in San Francisco (and online), and it was a big one. Satya Nadella's keynote alone ran over two and a half hours, and the Book of News is long enough that most people will only sk

Microsoft Ignite 2026 ran November 17–20 in San Francisco (and online), and it was a big one. Satya Nadella's keynote alone ran over two and a half hours, and the Book of News is long enough that most people will only skim the headlines.

So I went through the Book of News, the GitHub changelog, the Microsoft AI model pages and the early reactions, and sorted everything into categories a developer actually cares about. Each section has the status (GA / preview / Frontier-only), the numbers Microsoft published, and links for further reading.

Quick legend: GA = generally available, Preview = public/private preview, Frontier = Microsoft's early-access program for Copilot customers.

TL;DR

  • Copilot is now pitched as "the new operating system for work": Word, Excel and PowerPoint live inside the Copilot app (GA), plus Workspaces, Code and an always-on agent called Autopilot.
  • Microsoft's own models got serious: MAI Code 2 and MAI Code 2 Flash ship in GitHub Copilot, Microsoft-Decision-1.1 is GA, and OpenAI's GPT-6.2 Astra/Sol are GA in Foundry.
  • GitHub Copilot now works with Azure DevOps repos, drives your real signed-in browser, does local code reviews with secret detection, and can route work to local models on your PC.
  • Azure shipped Cobalt 200 Arm VMs (up to 50% faster than Cobalt 100) and HorizonDB, a distributed managed PostgreSQL (GA).
  • Agent 365 became the control plane for agents (50M+ agents, per Nadella) and gained cost management, because AI bills are now an IT problem.
  • Security got Microsoft Perception (GA), Defender runtime protection for local coding agents like Claude Code and Copilot CLI, and an AI Transparency Initiative.

The keynote in one tweet thread

Nadella posted his own summary thread on launch day:

// Detect dark theme var iframe = document.getElementById('tweet-2122793040058442116-245'); if (document.body.className.includes('dark-theme')) { iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2122793040058442116&theme=dark" }

He also opened the keynote by addressing the global Microsoft 365/Outlook outage from the week before. His words: "We have work to do." Keep that in mind, because reliability comes up again and again in the reactions below.

// Detect dark theme var iframe = document.getElementById('tweet-2122874904903397495-16'); if (document.body.className.includes('dark-theme')) { iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2122874904903397495&theme=dark" }

1. Copilot: rebuilt around Chat, Cowork, Code and Autopilot

The rebuilt Copilot is organized around four modes in one app instead of a pile of separate copilots.

Announcement What it is Status
Copilot Home Chat and Cowork in one place GA (rolling out worldwide)
Office in Copilot Word, Excel and PowerPoint inside the Copilot app, included with a Copilot license GA
Copilot Workspaces Shared team spaces where people and agents work toward the same goal with shared context (evolution of Copilot Notebooks, also shows up in Teams) Frontier
Copilot Code Build dashboards, trackers and apps in natural language; local sandbox on Windows, macOS coming Frontier, GA planned for December
Autopilot A long-running personal agent with its own identity and memory; colleagues can @mention your Autopilot in Teams Frontier
Copilot Managed Runtime Hosting for what people build in Copilot, with an optional Fabric data backend GA
Today (Brief + Feed) Morning briefing and priority feed Private preview
Copilot Cowork Auto effort levels with parallel subagents, Brand Kit, app building, Cowork in Edge GA + Preview

Why developers should care: "Copilot Code" and the Managed Runtime mean a lot of internal tools will now be built by non-developers. Someone will have to govern, debug and secure those apps, and that someone is probably you.

2. Models: MAI grows up, GPT-6.2 lands in Foundry

This was the most interesting part of the event for me. Microsoft is no longer just reselling OpenAI. It now ships its own model family and openly pitches them on price-performance rather than raw frontier capability.

What shipped

Model What it's for Status
MAI Code 2 Agentic coding model (2nd gen) GA in GitHub Copilot (model picker + Copilot Fusion)
MAI Code 2 Flash Cheaper, faster coding model GA in GitHub Copilot, limited preview in Foundry
Microsoft-Decision-1.1 Small model for routing, classification, verification and workflow control GA in Foundry, OpenRouter and Vercel
GPT-6.2 Astra & GPT-6.2 Sol (OpenAI) Frontier general models GA in Foundry
MAI-Image-2.7 + Flash Image generation Preview
MAI Cyber 2 Flash Security model Preview, only inside Codename MDASH
Meta Muse Wonder / Spark 1.3 / Ember Meta models Coming soon to Foundry
MAI Licensed Weights Model weights the customer controls Coming soon (Frontier Company engagements)

In total, Microsoft announced seven lower-cost MAI models spanning coding, cyber, image, voice and transcription, and Foundry's catalog now has 11,000+ models.

Benchmarks: what actually improved

These are Microsoft's own evaluations, so treat them as a reason to test, not as gospel.

MAI Code 2 (averaged across DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1):

Compared with MAI Code 2 performance MAI Code 2 cost
GPT-6.1 Sol ~98% ~40%
Claude Opus 5.5 ~96% ~5%
  • In blind frontend-app evaluations (258 annotators), developers preferred MAI Code 2's output over GPT-6 Sol 64% of the time.
  • MAI Code 2 Flash performs comparably to GPT-6 Luna and Claude Haiku 5.5 at roughly 65% and 20% of their cost respectively.
  • Caveat 1: the cost figures use a 50% promotional price.
  • Caveat 2: Microsoft itself says the models are "explicitly not yet tuned for harder problems." In other words, great for everyday enterprise coding, not your gnarliest refactor.

Microsoft-Decision-1.1 (the routing model):

  • Built on Gemma 4 26B A4B with image input.
  • Scores 68 on the Artificial Analysis Decision Index v2.1, up 6 points from the previous version.
  • Runs more than 2x faster than its predecessor.
  • More than 4,000 customers have already used it, and Microsoft uses it internally for routing, data labeling, personalization and content filtering.
  • Input tokens start at $0.042 per million.
  • This is the model that decides how much reasoning effort a Copilot task needs, and Epic is already using it across patient experience, clinical and back-office apps.

GPT-6.2 in Foundry (Global Standard, short context):

Model Input / 1M tokens Output / 1M tokens
GPT-6.2 Sol $2.00 $10.00
GPT-6.2 Astra $8.00 $40.00

Office-tuned MAI Work Flash also got numbers: on par with GPT-6 Luna for Excel and PowerPoint, close to GPT-6.1 Sol in Word with ~42% fewer input tokens per call, and up to 28% fewer tokens than GPT-6 Luna on Teams instruction-following and tool calling.

A real customer number: Alaska Airlines used a custom MAI model to lift maintenance-report classification accuracy from 81% to 91%.

The bigger story: Model routing is now a first-class product. Foundry's model router covers the top 50 Foundry models, Copilot Fusion routes coding tasks between cloud and local models, and Decision-1.1 picks the effort level. Microsoft's bet is that most tasks don't need the most expensive model, and that the platform should decide for you.

Further reading: Introducing MAI Code 2 and MAI Code 2 Flash

3. GitHub Copilot: the developer section

If you only read one section, read this one.

GitHub Copilot app

  • Azure DevOps repos (public preview): the Copilot app can now take ADO repos from clone to merge, not just GitHub repos. GitHub Enterprise Server instances are supported too.
  • Browser use (public preview): Copilot can fill forms, pull data from pages, and test and debug your web app inside the browser you're already signed in to (Chromium-based: Chrome, Edge).

Copilot code review

  • Local reviews in the CLI and the app (review before you push).
  • Secret detection during review.
  • Adaptive review effort, so small diffs don't burn a big review.
  • Cross-repository context and repository insights.

Copilot Fusion + local models

  • Copilot Fusion routes each coding task to the right model and can now route work to local models on your developer PC (preview, Windows). That's a big deal for cost and for teams that can't send everything to the cloud.

Microsoft Copilot SDK 2.0 (GA)

  • The same agent harness Microsoft and GitHub run in production, rebuilt on a native Rust runtime, available in six languages, built for multi-tenant, high-scale agent deployments.

Governance and evaluation

  • GitHub Copilot Evals: test whether a customization (instructions, prompts, skills) actually helps before you roll it out.
  • Agent governance rulebooks and bring-your-own-model in the CLI and IDEs: GA.
  • Copilot activity streams into Agent 365 (prompts, responses, tool calls), with risk signals in Purview and Defender.
  • Microsoft Entra will assign an Agent ID to discovered local agents, starting with GitHub Copilot CLI and Claude Code.

JetBrains

Copilot for JetBrains gets remote access (continue a chat in VS Code or the Copilot app), on-device voice dictation, browser-use controls, historical diffs and Copilot Fusion pinned in the model picker.

Further reading:

4. Agent 365: managing agents (and their bill)

Nadella said there are already more than 50 million agents in Agent 365, and described the new analytics as an ERP system for AI usage.

Announcement What it does Status
Agent 365 Cost Management Spend drivers, policies, limits and alerts for usage-based AI spend Core in preview; forecasting via Frontier
Agent 365 Business Analytics Ties AI spend to outcomes (success rate, satisfaction, assisted hours) Frontier
Dynamic Agent Risk Combines runtime, posture and blast-radius signals into a risk level (partners: Check Point, Cyera, Darktrace, Zenity) Preview
Plugin and skill governance Central control over plugins/skills GA
MCP governance for Azure AI Gateway and LiteLLM Govern MCP servers Preview
Local and shadow agent governance Find and govern agents running on endpoints Preview

5. Microsoft IQ and Foundry

Microsoft IQ (GA, no new SKU) is the enterprise context layer under Copilot. It combines Work IQ, Fabric IQ, Foundry IQ and Web IQ, with 350+ third-party connectors.

  • Fabric IQ connectors for ChatGPT and Claude (preview): your Fabric data as context in non-Microsoft assistants.
  • Work IQ integrations with OpenAI and Anthropic are coming soon.
  • Web IQ is GA for all Azure customers, starting at $1 per 1,000 calls.
  • Foundry continuous improvement: multi-turn evaluation, insights and Agent Optimizer are GA; human annotation and managed training are in preview.
  • Copilotforce: a Microsoft and Salesforce partnership bringing Salesforce data and actions into Copilot, Teams and Fabric (Salesforce in Copilot has 38+ prebuilt sales skills via authenticated Salesforce MCP).

6. Azure: compute, data and infrastructure

Compute

  • Azure Cobalt 200 VMs (GA): Microsoft's next-gen Arm VMs with up to 50% better CPU performance than Cobalt 100, consistent-speed and boost modes, heading to 24 regions by year-end, plus new Mpsv4 (large in-memory DBs) and Lpsv5 (up to 23 TB local storage) SKUs.
  • AKS for agents (preview): deploy Foundry models and agents to AKS, Agent Sandbox extension for AKS with VM-level isolation, Agent Grid, and Flex nodes on your own servers.
  • Confidential computing (preview): VR72 multi-GPU confidential computing, Confidential Enclaves, AKS Kata CoCo containers, Confidential Clean Room.

Data

  • Azure HorizonDB (GA): distributed managed PostgreSQL. Microsoft claims up to 8x faster OLTP than self-managed PostgreSQL (its own HammerDB benchmark) and scaling to 128 TB. Entra auth, private endpoints, CMK and multi-zone replication included. Fabric HorizonDB (serverless, pause/resume) is in preview.
  • Cosmos DB: Rust SDK, Shell for VS Code and portal, semantic reranker, Azure Backup all GA.
  • Azure Managed Redis: Redis 8 + an agent kit in preview.
  • Azure SQL Database container for local development (preview).
  • PostgreSQL: Oracle-to-PostgreSQL migration in DMS (preview).

Ops and migration

  • Azure Catalyst (GA): AI-powered discovery, planning and migration, including the Catalyst agent (formerly the migration agent).
  • Azure SRE Agent: customer-managed keys, Entra Agent ID, Teams integration and multicloud investigations (preview).
  • Chaos Studio workspaces and AKS extension: GA.

Further reading: Azure Cobalt 200 VMs now available

7. Security

  • Microsoft Perception (GA): teams of red, blue and green security agents running playbooks in a closed loop, with humans approving high-impact actions.

// Detect dark theme var iframe = document.getElementById('tweet-2122793047448862890-779'); if (document.body.className.includes('dark-theme')) { iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2122793047448862890&theme=dark" }

  • Defender runtime protection for local AI agents (GA, Defender for Endpoint P2): covers Claude Code, GitHub Copilot CLI and app, Claude Desktop and VS Code extensions, with network-based protection for OpenClaw, Ollama Desktop and ChatGPT.
  • Codename MDASH (preview): agentic code security using 100+ agents in four stages, now with repository sharding and MAI Cyber model support.
  • Vulnerability shielding (preview): temporary controls on vulnerable systems during active attacks.
  • Integrated SOC: core GA, full SIEM + 90-day Defender retention in preview.
  • AI Transparency Initiative: launched with 40+ organizations to make model and agent behavior more verifiable. Nadella's framing was that we can see what models touch, but not what happens inside them.

8. Windows and Windows Server

  • Windows Server 2028 (Insider preview): hardened Server Core, post-quantum cryptography, Trusted Launch, NVMe over Fabrics, booting from ReFS, and local AI via Windows ML, Foundry Local, vLLM and llama.cpp. New hardware support includes AMD MI350P and NVIDIA GB300 Desktop.
  • Hybrid intelligence on Windows: GitHub Copilot Fusion can tap local models on developer PCs; local context, actions and models are coming to Copilot on Copilot+ PCs.
  • Direct Stream for Windows 365.
  • Edge for Business: skills recorder with Cowork and managed agent browsing (preview).

Further reading: Building Windows for the next era of work

9. Identity, Security & Windows: the rest of the list

The second half of the Book of News is easy to miss, but it has some of the most practical items for anyone running AI agents at work.

Entra (identity for agents)

  • Microsoft MCP Server for Enterprise (GA): natural-language, permission-based access to Entra data.
  • MCP Firewall (preview): control access to MCP servers, tools, resources, prompts, protocols and transports.
  • Agent ID for discovered local agents (starting with GitHub Copilot CLI and Claude Code) and for Autopilot agents (preview).
  • Global Secure Access: network visibility and control over local agents and the MCP servers they talk to (preview); agentic traffic acquisition is GA.
  • Also in preview: Conditional Access policy plans, high-assurance credentials, Entra Backup and Recovery Premium (30-day retention), and cross-tenant configuration management.

Intune: governing local AI agents (preview)

  • Discovers agents on managed devices through Agent 365.
  • Marks devices running unapproved agents as noncompliant and blocks them with Conditional Access.
  • Sets boundaries for approved agents using Microsoft Execution Containers.

More on Microsoft Perception

  • The Triage Agent covers about 80% of alerts in a typical Defender queue; the Investigation Agent follows attacks across domains.
  • Third-party coverage now includes AWS, Salesforce and Okta for triage, investigation and detection authoring.
  • Web App Pen Test Agent (preview): validates exploitable weaknesses with evidence and remediation guidance.
  • A "detect emerging threats" playbook writes tenant-aware KQL detections and tests them against historical data.

Purview: protecting data from Copilot and agents (preview)

  • Copilot DLP blocks Copilot from referencing sensitive files even when they aren't labeled, and extends prompt protection to attachments, linked files and text.
  • Data security now extends to agents running on Windows endpoints.

Windows and Windows 365

  • Copilot on Windows (preview): plain-language local file management (find, organize, move, copy, rename) and PC diagnostics that explain issues and suggest fixes, reachable from the new Windows Search.
  • Cloud PCs for local AI models (GA): up to 32 vCPUs and GPU-enabled Cloud PCs with Foundry Local.
  • Direct Stream with NVIDIA: up to two 4K streams at 60 fps for GPU Cloud PCs (preview early next year, select regions).
  • Process containers for AI agents: protected workspaces with limits on communication, network access and identity (preview); a learning/permissive mode that observes agent behavior before enforcing rules is GA.
  • Windows 365 for Agents is now eligible for Copilot Credits, with a Foundry MCP server coming early next year. OpenAI and Meta are already building on Microsoft Execution Containers.
  • Also new: Windows Endpoint Security Platform (GA), Windows 365 admin assistant in Intune and Windows MAM (preview).

Smaller Azure and Foundry items

  • Foundry IQ Serverless: usage-based pricing with a free monthly allowance, GA planned for December.
  • Azure Copilot deployment agent (GA): goes from requirements to deployed workloads.

Why this matters for developers: if your team runs Claude Code, Copilot CLI or other local agents on company laptops, IT now has the tools to discover them, give them identities, firewall their MCP traffic, and block unapproved ones. Expect "which agents are allowed on my machine?" to become a real policy conversation.

What people are saying

Not everyone is cheering. Here's a cross-section of the reactions:

  • Satya Nadella (Microsoft CEO) called Copilot "the new operating system for work" and argued it's time to move beyond single-player AI toward a multiplayer form factor. He also told reporters he's "AGI-pilled."
  • Carolina Milanesi (Creative Strategies) argued that cost was the real headline. She notes that cost came up in almost every keynote segment, and that Microsoft gives customers better tools to watch and cap spend while pricing stays metered and largely in Microsoft's hands. She also points out that the keynote demos alone went past the default per-user monthly credit limit, so bills will grow as companies raise those limits.
  • Chris Hoffman (PCMag) was sharper. His take: Microsoft is asking customers to bet their business on Microsoft 365 days after a massive outage, and none of the announcements matter if the service goes down. He also noticed that Microsoft 365, not Windows, is now the center of Microsoft's AI strategy.
  • DevOps.com framed the coding announcements as Microsoft making the case for "good enough" coding models: cheaper, task-specific models plus routing, sandboxes and governance, treating AI coding as enterprise infrastructure rather than just a developer tool.
  • Cloud Wars read the keynote as Microsoft wanting to own the context layer, not just the Copilot UI.
  • WindowsForum pointed out a gap: Nadella called MAI Cyber 2 Flash Microsoft's most capable security model, yet it's only in preview inside MDASH.

My own take: the most developer-relevant shift is routing + cheaper in-house models + governance. If MAI Code 2 really delivers ~96–98% of frontier quality for a fraction of the price on everyday tasks, the "which model?" question becomes a platform default rather than a developer choice. Whether that's good depends on how much you trust the router.

Status cheat-sheet

Generally available now Preview / Frontier / coming soon
Office in Copilot, Copilot Home, Managed Runtime Copilot Workspaces, Autopilot, Copilot Code (GA Dec)
MAI Code 2 + Flash (GitHub Copilot) MAI Code 2 Flash in Foundry, MAI-Image-2.7, MAI Cyber 2 Flash
Microsoft-Decision-1.1, GPT-6.2 Astra/Sol Meta Muse models, MAI Licensed Weights
Copilot SDK 2.0, Copilot BYOM + governance rulebooks Copilot app ADO support + browser use, Fusion local models
Microsoft IQ, Web IQ Fabric IQ for ChatGPT/Claude, Work IQ for OpenAI/Anthropic
Cobalt 200 VMs, HorizonDB, Azure Catalyst Fabric HorizonDB, AKS Agent Sandbox, confidential GPU
Microsoft Perception, Defender for local agents Agent 365 Cost Mgmt (core), Dynamic Agent Risk, MDASH
Entra MCP Server for Enterprise, Cloud PCs for local AI MCP Firewall, Intune local agent governance, Copilot DLP
Agent learning mode, Windows Endpoint Security Platform Copilot on Windows, process containers, Foundry IQ Serverless
— Windows Server 2028 (Insider)

Further reading

Official

Analysis

Which announcement matters most to you: MAI Code 2 in Copilot, local models via Fusion, or HorizonDB? Let me know in the comments. 👇

📰 Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.