Satya Nadella opened Ignite 2026 by competing on three fronts at once: frontier capability, price-performance, and verifiable AI. Microsoft is not conceding the first to win the second. It is arguing that all three matter, and the partner reading is that the economics and the governance each create a service opportunity.
Key Takeaways
- Microsoft AI announced seven models with different availability paths, not one launch. Its accompanying article reports MAI Code 2 at 98 percent of GPT-6.1 Sol's performance for 40 percent of the price, averaged across four coding benchmarks.
- Autopilot is the fourth Copilot form factor, alongside Chat, Cowork and Code: an autonomous agent that keeps moving work forward 24x7 wherever the team collaborates.
- Microsoft IQ is positioned as a platform for an organisation's own intelligence, naming Work IQ, Fabric IQ and Web IQ, with hundreds of new connectors and business plugins that can be shared and inherited.
- Agent 365 Business Analytics is Microsoft's attempt to connect AI spend to outcomes. Nadella called it an ERP system for AI usage, and over 50 million agents are already in the control plane.
- Microsoft Perception is generally available, with red, blue and green agents orchestrated through playbooks.
- The AI Transparency Initiative with 40 organisations is a plan, not a shipped feature, and Microsoft says it will incorporate the standards into MAI models, Foundry and Copilot.
The models, and where each number comes from
Microsoft's own model announcement is the source for the ratios below. Treat the availability status of each one separately.
- MAI Code 2: 98 percent of GPT-6.1 Sol coding quality at about 40 percent of the price, averaged over four benchmarks (DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1). Against Claude Opus 5.5 the article reports 96 percent at 5 percent of the price. Nadella described Code 2 as at the frontier of cost-performance for everyday enterprise coding tasks, and said it is generally available today in GitHub Copilot, both in the model picker and as part of Copilot Fusion, and available in Foundry.
- MAI Code 2 Flash: comparable performance to GPT-6 Luna and Claude Haiku 5.5 at 65 percent and 20 percent of their respective prices, across three benchmarks (DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas). In GitHub Copilot, and in a limited access preview in Foundry.
- MAI Cyber 2 Flash: Microsoft reports performance ahead of frontier models at a fraction of the cost per task. Keep two numbers apart: the claim that it matches GPT-6 Astra quality at under 2 percent of the cost per task on CyberGymE2E is a different measurement from the claim that it is 86 percent cheaper on output tokens than Cyber 1.1 Flash.
- MAI Image 2.7: reported first for image editing on Artificial Analysis, at one fifth of the price of GPT Image 2.6 Sunburst (max).
- MAI Realtime 1: in Microsoft's evaluations, ahead of GPT-Realtime 2.1 across external and Dynamics 365 benchmarks, starting at one US cent per audio minute.
- MAI Transcribe 2: powers Dragon Copilot Transcribe across 10 million patient encounters a month, with multilingual error rates halved.
- Microsoft-Decision-1.1: generically available. Nadella described it as more than 10 times faster and 200 times less expensive than using an LLM, without naming the comparison model. Microsoft's separate Decision article reports 50x faster and 79x cheaper than GPT-Sol 6.1. Those are two different comparisons and should not be merged. The article also lists routing, classification, prioritisation and verification as intended uses.
Microsoft also flagged two more model announcements: MAI Licensed Weights will let trusted customers access Microsoft AI model weights through Frontier Co, and Foundry is adding GPT 6.2 alongside Meta's latest models.
One detail worth keeping: the coding models are explicitly not yet tuned for fundamental science research or very long horizon agentic work. Not tuned is not the same as incapable, and Microsoft said it themselves.
Copilot is the new OS for work, and Autopilot is the 24x7 half
Nadella framed Copilot as the new OS for work and named four form factors in one breath: Chat, Cowork, Code and Autopilot.
- Home is where you start, for a quick answer or a multi-step task.
- Code gives one person the power of a SaaS factory, building apps and dashboards inside the tenant.
- Cowork is the shared work surface.
- Autopilot is, in his words, an autonomous agent that keeps moving your work forward 24x7 in all the places where you and your team collaborate.
Underneath all of it sits Work IQ, which he called the most important database for any organization: the understanding of people, relationships, documents, emails, meetings, chats and projects that make up the fabric of work.
Two things follow for a partner, and both are our reading rather than Microsoft's claims.
First, Autopilot changes the consumption curve. An agent working around the clock in every collaboration surface does not consume in the pattern a seat licence assumes. If Cowork already made spending policies necessary, a 24x7 agent makes them more urgent. Every Autopilot rollout is a cap, an owner and an alarm waiting to be set.
Second, the spend-to-outcome link is now a Microsoft message. Nadella introduced Agent 365 Business Analytics as an ERP system for AI usage: see what drives spend, connect spend to outcomes, optimise ROI. Over 50 million agents are already in Agent 365, the control plane that extends existing governance, identity, security and management to agents. That is a stronger basis for a partner economics conversation than speculation about repricing.
Microsoft is also extending Copilot by role, for sales, legal, finance and service, with new business application plugins starting with Dynamics 365, and a partnership to bring Salesforce into Copilot. No licensing or general availability terms were given for these.
Infrastructure rebuilt for agents
GitHub is rebuilding its core. Agent-driven pushes, commits, actions and pull requests have grown exponentially, so the git architecture is being reworked, storage and compute decoupled, and the push critical path shortened. The test number from the keynote: 5,000 agents on a single repository sustained 7,200 writes per second, which Microsoft described as 180 times current throughput.
Azure has more than 500 data centres and 80 regions, with community-first AI infrastructure commitments attached. On sustainability, Microsoft said its annual global electricity consumption is matched with renewables, that 92 percent of servers and components are reused or recycled, and that cooling is water-conserving. Sovereign cloud options give customers control over where workloads run.
The agent primitives you will be asked about
Microsoft described three new building blocks for running agents:
- Micro-VMs that start in under a second with full sandbox isolation, able to run inside confidential computing enclaves.
- An agent grid adding identity and policy-aware networking, controlling what each agent can reach and which agents it may talk to.
- A context cache that accelerates model loading, with the stack wrappable in AKS, AKS Flex, or hosted agents in Foundry.
The agent grid is the commercially interesting one, in our view. It is the difference between an agent that has network access and an agent that has the specific access it was granted, to the specific things it was granted. That is a control security teams have been asking for.
Windows: unmetered intelligence at the edge
Windows is extending AI to the edge with local models, described on stage as unmetered intelligence: last year's frontier models running on laptops, last quarter's on desktops.
A device that runs a capable model locally does not meter every token against a cloud bill. For regulated customers and for anyone with a large field workforce, that changes the shape of an AI deployment conversation, and it changes the device refresh conversation with it.
Microsoft IQ: the layer you will actually sell
Nadella described IQ as a way to help every organisation build, retain and continuously improve its own intelligence, by externalising the system underneath Copilot as a first-class platform. He named Work IQ, Fabric IQ and Web IQ, bringing together data across Fabric, Foundry and Microsoft 365, plus third-party context, with hundreds of new connectors and context that manifests as business plugins which can be shared and inherited across the organisation.
The platform provides connectors and governance capabilities. Customers still need to configure integrations, permissions and reuse policies, and we see that implementation work as the partner opportunity rather than a licence line.
Copilot Workspaces adds the multiplayer form factor, bringing people and their agents together around a common goal with shared context.
The data layer: HorizonDB plus Microsoft's Ignite news
This section combines Nadella's HorizonDB announcement with Microsoft's accompanying Ignite news on Cosmos DB and Azure DocumentDB.
Azure HorizonDB is generally available, with Fabric HorizonDB in public preview. Microsoft reports up to 8x faster OLTP than self-managed PostgreSQL on an internal benchmark, databases up to 128 TB, zone redundancy by default, sub-millisecond commit latency and failover under five seconds. Fabric HorizonDB provisions in seconds, pauses compute after five minutes idle and resumes on the next query, and it is the default database for Fabric apps. Security capabilities include Entra ID authentication, private endpoints and customer-managed keys. These are capabilities, not a compliance verdict. Suitability has to be assessed against the customer's own requirements, and the 8x figure should be validated against the customer's workload before it becomes a migration case.
Cosmos DB carries two stories worth separating. Microsoft's article reports that OpenAI's Habitat platform supports products used by more than a billion people each week, at over 70 million requests per second and more than 500 petabytes. The 6x CPU and 15x memory gains belong to OpenAI's own rearchitecture of Habitat from Python to Rust, not to a general runtime comparison. Serverless and multi-tenancy capabilities, including throughput buckets, are what Microsoft says let OpenAI share a database across hundreds of workloads and deploy and scale new workloads to 1.2 billion users in a day rather than months. Separately, LSEG moved several legacy Oracle databases of record to Cosmos DB, and Microsoft now has a relational-to-Cosmos DB modernization assistant in preview and a Cosmos DB Agent Kit at general availability.
Azure DocumentDB is Microsoft's fully managed, MongoDB-compatible service, built on the open-source DocumentDB project governed by the Linux Foundation. That governance applies to the open-source project, not to the Azure service. The VS Code extension and near instant clusters in the M10 to M30 tiers are generally available, and full-text search with BM25 relevance is in public preview alongside existing vector search.
Our judgement is that the database announcements deserve more partner attention than any single model release, because PostgreSQL, Oracle and MongoDB workloads each carry migration and validation work that does not arrive with a licence key.
The transparency bet
The most consequential announcement for anyone selling security or governance is not a model.
Nadella put the problem sharply: today we can see almost everything models touch, but we cannot see what happens inside the model, and what a model tells us about itself is not independent evidence. He called these nested black boxes unacceptable.
The AI Transparency Initiative follows, with three parts: expanded collaboration in NIST's Agent Standards Initiative, an ecosystem approach with 40 organisations working to define standards that apply regardless of model or provider, and an intent to incorporate the same standards into MAI models, Foundry and Copilot. That third part is a plan, and the announcement gives no delivery date.
One line worth keeping in a deck, from Microsoft's press release rather than the stage:
"Blaming the AI isn't an answer, and 'trust the model' isn't a control."Microsoft Corp., press release via PR Newswire, November 17, 2026.
What this may do to the money, and what we do not know
This section is our assessment. Microsoft did not announce pricing or margin changes.
Microsoft's own models may give it more flexibility in inference economics, since Code 2 is already inside GitHub Copilot. What that does to Copilot costs, list prices or reseller margins has not been disclosed, in the keynote or in the accompanying material. Partners can stress-test their margin assumptions without treating repricing as announced.
What was announced is a control plane. Agent 365, Perception and the transparency work all answer the same customer question: what did the agent do, and can you prove it. Microsoft is shipping the controls. Turning them into an operating model for a 200 person customer is still open work, and that is where we see the billable opportunity.
Our view is also that cheaper code generation increases the value of testing and verification, because generating code stops being the scarce step.
What is still missing
Nadella's keynote did not include Copilot Credits pricing, an Agent 365 price list, or dates for the transparency standards. We could not access the Book of News during our review. Several figures in this piece, including the GitHub throughput test, the Maia 300 claim and the sustainability numbers, come from the keynote itself and had no published Microsoft page at the time of writing. Treat every figure as sourced from Microsoft on stage or in the article cited beside it, and treat every commercial implication as ours.
Partner Play
Monday morning, in this order:
- Do not lead with the models. Price-performance is evolving quickly, and today's model price is not the whole customer story.
- Map Microsoft IQ to one customer. Pick the account with the messiest data estate and price the connector, permission and reuse-policy work.
- Put HorizonDB on the PostgreSQL estate. Any customer running self-managed PostgreSQL for an AI or RAG workload is a benchmark conversation, provided you validate the 8x figure against their workload first.
- Turn Perception and Agent 365 into an assessment. Agent inventory, identity review, spending policy, evidence trail. Sell it before the first agent goes into production, not after.
- Build the spend-to-outcome report. Agent 365 Business Analytics exists, but somebody has to define what an outcome is for this customer and report against it monthly.
- Re-examine your Copilot margin assumptions without assuming repricing. Scenario-test them, and say plainly that Microsoft has not announced a change.
- Follow the Book of News for a consolidated overview, and use the official product announcements to verify details and availability.
Sources
- Microsoft AI: Seven new world class models at scalable price-performance - Official Microsoft Blog: Ignite 2026, building the frontier ecosystem - Microsoft press release via PR Newswire: AI Transparency Initiative - Microsoft Command Line: Microsoft-Decision-1.1 is now generally available - Microsoft Tech Community: What's new with Microsoft Perception at Ignite 2026 - Microsoft Tech Community: What's new in Work IQ at Ignite 2026 - Microsoft Tech Community: HorizonDB, PostgreSQL for enterprises, developers and agents on Azure and Fabric - Microsoft Dev Blog: What's new in Azure Cosmos DB at Microsoft Ignite 2026 - Microsoft Dev Blog: What's new in Azure DocumentDB at Microsoft Ignite 2026
Deep Microsoft licensing, incentive, and Partner Center guidance for resellers is published on Cloud Factory.
Ready to grow your Microsoft practice? Explore Cloud Factory's partner platform →
Looking for your local Cloud Factory team? Find your country's contact →