Microsoft used its Ignite conference in San Francisco on November 17, 2026 to announce six new in-house AI models for business services. The launch gives the company a lineup of its own for coding, cyber defense, image work, audio and decision-making. It also marks the clearest step yet away from a stack that leaned on OpenAI for years.
The details matter more than the headline count. Microsoft named the MAI Code 2, MAI Code 2 Flash, MAI Image 2.7, MAI Image 2.7 Flash and MAI Cyber 2 Flash models, plus MAI Realtime 1, MAI Transcribe 2 and Microsoft-Decision-1.1. That is more than six products, and the sources do not agree on which six make the official cut. This analysis sorts out what the record confirms, what rests on Microsoft’s own benchmark claims, and what remains unpriced.
Timing adds pressure. India Today reports that OpenAI revealed GPT-6.2 Astra and GPT-6.2 Sol as Microsoft took the stage, so the price war now runs on both sides of a partnership that once looked simple. Our breakdown of GPT-6.2 Astra and Claude at price parity covers the rival side of that contest. This piece focuses on what Microsoft shipped, what enterprises can license, and what the numbers would mean if they hold up under independent testing.
Don't miss new tech stories on Google
Add Shattered once in the Google app and our stories show up in your news suggestions.
[Add now](https://www.google.com/preferences/source?hl=en&q=https%3A%2F%2Fshattered.io%2F)
## The Six-Versus-Seven Count: What the Sources Confirm
Ticker News, India Today and several aggregators put the number at six. Other coverage counts seven. Microsoft’s own Ignite page lists Microsoft-Decision-1.1 in place of the MAI Realtime 1 and MAI Transcribe 2 pair, which means no single source in the available record shows a clean six-model list. Readers should treat “six” as the headline shorthand and the named products as the real inventory.
The table below lays out every product Microsoft has named, what it targets, and the status each source supports. Where the sources stay silent, the table says so instead of guessing.
| Model | Focus area | What the sources confirm |
|---|---|---|
| MAI Code 2 | Long-horizon, multi-file coding | Reasons across an entire repository, planned for GitHub Copilot |
| MAI Code 2 Flash | Faster, cheaper coding | Named by Microsoft, pricing not published |
| MAI Image 2.7 | Image generation and editing | Microsoft says it ranks first in image editing |
| MAI Image 2.7 Flash | Lighter image model | Named by Microsoft, no specs in the record |
| MAI Cyber 2 Flash | Cyber defense | 12 billion active parameters, per Microsoft |
| MAI Realtime 1 | Fast voice processing | Named in the announcement, no specs published |
| MAI Transcribe 2 | Transcription, including healthcare use | Named in the announcement, no specs published |
| Microsoft-Decision-1.1 | Routing, classification, workflow control | Microsoft’s page calls it generally available |
Eight rows, one launch, and only three or four of them carry hard technical detail. That gap defines the first week of coverage. The announcement tells buyers where Microsoft wants them to look, but it does not yet give procurement teams a price sheet.
MAI Code 2 and Code 2 Flash: The Coding Pair #
MAI Code 2 is the flagship of the batch. Microsoft built it for complex, long-horizon, multi-file coding tasks and says it can reason across an entire repository. That framing targets the hardest agentic work, where a model must hold a large codebase in view and make coordinated edits instead of completing a single function.
MAI Code 2 Flash takes the opposite role. India Today describes it as the smaller, faster version, which suits everyday work where speed and cost count more than peak reasoning depth. Splitting the family this way mirrors how teams actually use coding assistants. A few hard problems need the large model, while thousands of small completions and edits do not.
Microsoft plans to ship MAI Code 2 inside GitHub Copilot, according to the reports. That distribution channel may matter more than any benchmark. Developers who already pay for Copilot would reach the model without a new contract, a new SDK or a new security review.
The headline performance numbers come from Microsoft, not from an outside lab. The company says MAI Code 2 reaches 98% of GPT-6.1 Sol’s performance at 40% of the cost, and 96% of Claude Opus 5.5’s performance at 5% of its cost. Those figures are reported claims, and no independent test in the available record confirms them. The comparison section below breaks them down.
MAI Image 2.7: Editing Rank and Preview Status #
MAI Image 2.7 and its Flash sibling cover image generation and editing. Microsoft says MAI Image 2.7 ranks first in image editing, which is a specific and checkable claim. Image leaderboards update often, so the ranking can change within weeks, and readers should look at the leaderboard itself instead of relying on a launch-day screenshot.
Editing is the right place to plant a flag. Generating a fresh picture from a prompt has become a commodity feature. Editing an existing marketing asset without breaking brand colors, faces or layout is the harder enterprise job, and it is where agencies and in-house design teams spend real hours.
The sources do not publish prices for either image model, and the available material does not detail the Flash version beyond its name and its place in the family. Anyone planning a budget around image workloads should wait for Microsoft’s price list. Expect the Flash tier to serve high-volume, lower-stakes jobs such as product thumbnails and social variants, though that reading is an inference from the naming pattern across the lineup.
MAI Cyber 2 Flash and the 12 Billion Parameter Figure #
MAI Cyber 2 Flash is the one model where Microsoft gave a concrete size. It runs on 12 billion active parameters, according to the company. That puts it in the small-model class, well below the frontier systems that headline most launches.
“Active” parameters is a term worth pausing on. It usually signals a design where only part of the full network fires for any given token, which cuts the compute cost per request. Microsoft has not published the total parameter count in the material reviewed here, so readers cannot judge the full architecture.
India Today reports that Microsoft says the model beats OpenAI’s GPT-6 Astra on a major security test while costing under 2% per task. As with the coding numbers, that comes from the vendor. Security teams should ask which test, which version and which tasks before they treat it as a buying signal. A cyber model that scores well on a benchmark but misses real attacker behavior would cost far more than any token savings.
Defenders have a practical reason to care about small models. A 12 billion parameter system can run closer to the data, respond faster and cost less per alert, which fits a security operations center that processes huge volumes of logs. Our AI and machine learning coverage tracks how other vendors approach the same trade-off.
Realtime, Transcribe and Decision-1.1: The Plumbing Models #
Three of the named products do less glamorous work. They handle voice, transcription and traffic control between models. None of them will trend on social media, yet together they may do more to cut an enterprise AI bill than any flagship.
Microsoft-Decision-1.1: routing inside applications
Microsoft describes Microsoft-Decision-1.1 as a model for routing, classification, prioritization, verification and workflow control, and its official page lists it as generally available. In plain terms, it decides which model or tool should handle a request. A router that sends easy queries to a cheap model and hard ones to an expensive model can cut spend without any single model getting better.
MAI Realtime 1 and MAI Transcribe 2: voice and healthcare audio
India Today says MAI Realtime 1 handles fast voice processing and MAI Transcribe 2 targets healthcare transcription. The sources give no latency figures, no word-error rates and no language lists. Without those numbers, buyers cannot compare either model with existing speech services, so these two remain announcements rather than evaluated products.
Licensed Weights Inside Azure: How the Enterprise Offer Works #
The business model may be the most interesting part of the launch. Microsoft lets enterprise customers license its model weights, customize the models with proprietary data and run them inside Azure. The customer gets a model tuned to its own material, and Microsoft keeps the deployment on its own cloud.
That differs from a normal hosted API. With a shared endpoint, every customer calls the same model and the vendor controls updates. With licensed weights, a company holds a copy it can adapt, which gives it more control over behavior and data handling. India Today reports Suleyman said the approach should feel similar to open-source software, with tighter control and security over a company’s information.
The catch is the deployment rule. Customers must run the weights within Azure, so the arrangement does not let a bank or hospital move the model to a rival cloud or an on-premises rack. For Microsoft, that keeps compute revenue in its own data centers even when the model itself ships as a license. Azure buyers already see the company pushing silicon and VM gains, as in our report on Azure Cobalt 200 VMs reaching general availability.
Licensing terms, fees and the list of eligible models did not appear in the sources reviewed. Microsoft’s Ignite Book of News is the place to watch for those details as the company fills in the program.
Microsoft’s Price-Performance Claims, Line by Line #
Every competitive number in this launch comes from Microsoft. India Today relays them, and the company’s own keynote supports the first pair. The table sorts them so readers can see which rival each model targets and what Microsoft says it delivers.
| Microsoft model | Rival named | Performance claim | Cost claim |
|---|---|---|---|
| MAI Code 2 | GPT-6.1 Sol | 98% of performance | 40% of cost |
| MAI Code 2 | Claude Opus 5.5 | 96% of performance | 5% of cost |
| MAI Code 2 Flash | GPT-6 Luna | Matches | 65% of price |
| MAI Code 2 Flash | Claude Haiku 5.5 | Matches | 20% of cost |
| MAI Cyber 2 Flash | GPT-6 Astra | Beats it on a major security test | Under 2% per task |
| MAI Image 2.7 | Image-editing field | Ranks first in editing | No price published |
Read the Code 2 rows carefully. A model that delivers 96% of Claude Opus 5.5 at 5% of the cost would reshape coding economics if the claim holds on a team’s real work. Averages hide variance, though. A model can score 96% across four benchmarks and still fail on the one repository an enterprise cares about.
The Flash comparisons say “matches” instead of giving a percentage, and India Today attributes those figures to Microsoft. Treat all six rows as marketing until an outside lab reruns them. Independent leaderboards usually take days or weeks to include new models, so the first credible checks should land before the end of the year.
What Suleyman Said on Stage #
Mustafa Suleyman, CEO of Microsoft AI, carried the pitch. His remarks fall into three themes: price, ownership and routing. On price, he told the audience, as Yahoo Finance reported: “Our commitment is to give you the very best price performance out there, so that you can really scale your agents.” The sentence sets the terms of the whole launch. Microsoft is selling cost per task, not raw capability.
On ownership, his keynote text on the Microsoft AI site reads: “Your core assets in this new platform shift are your IP and your workflows, the years of commercial data you’ve assembled, essentially everything that makes your company what it is.” That line sells the licensed-weights idea. It tells a chief information officer that a customized model carries a company’s accumulated knowledge, so the company should control it.
The coding claim came in the same keynote. Suleyman said: “Averaged over four widely used benchmarks that target the everyday enterprise coding tasks we see in GitHub Copilot, MAI Code 2 gives you 98% of the performance of GPT 6.1 Sol at just 40% of the cost.” Note the wording. He anchors the test to everyday enterprise coding, not to frontier research problems, which fits India Today’s report that the specialized models give up breadth for price.
On orchestration, he said in the same keynote: “So Decision 1.1 provides model routing integrated right into your applications.” Routing may sound like a footnote, yet it ties the lineup together. A company that adopts Decision 1.1 can mix Microsoft models with others behind one interface, and Microsoft keeps its place in the middle of every request.
Microsoft Versus OpenAI, Anthropic and Google: The Competitive Picture #
Ticker News says the new models put Microsoft in direct competition with OpenAI, Anthropic and Google. Axios framed the launch the same way, reporting that Microsoft aims to challenge OpenAI and Anthropic with cheaper, specialized AI models. The strategy is narrow by design. Instead of one giant model for every question, Microsoft ships smaller models that focus on a single job.
Specialization as a pricing weapon
India Today reports Suleyman’s explanation that the price drops come from packing more knowledge into a smaller space. A model that only writes code does not carry the weight of advanced math or scientific research, so it costs less to serve. The trade-off is plain. A coding model will not solve a hard physics problem, and Microsoft does not claim it will.
Ticker News adds that Suleyman confirmed work on a general-purpose model built to frontier standards. Details on its release and pricing remain unannounced. Until it arrives, Microsoft competes on price in narrow lanes while its rivals compete on breadth.
Rivals have their own moves. Anthropic just priced its IPO, as covered in our report on the Anthropic IPO at $83 a share, which hands it fresh capital to answer price cuts. OpenAI answered the same week with its GPT-6.2 releases, and India Today says it made Astra 20% cheaper than its previous version. Microsoft now sells against the companies whose models it also hosts.
GitHub Copilot as the Distribution Channel #
Microsoft owns GitHub Copilot, and it plans to put MAI Code 2 inside it. That gives the company something rivals lack: a built-in audience of developers who see model choices every day. If Copilot routes routine requests to a cheaper in-house model, Microsoft cuts its own cost of serving each request.
The economics run in two directions. Lower per-request costs help Microsoft’s margins, and they let it hold subscription prices steady while usage grows. For customers, the question is whether the cheaper model matches the quality they expect. Teams that track spend per active developer, as in our Cursor versus Copilot seat audit, can test the new option against their own usage data.
Reports describe MAI Code 2 as planned for Copilot, and the sources reviewed here do not give a rollout date, plan tier or regional list. Until Microsoft publishes those, no developer should assume the model appears in a given plan on a given day. Watch the GitHub changelog for the first concrete dates.
Market Impact: What Changes for Enterprise AI Budgets #
Pricing pressure is the first effect. When a hyperscaler says its coding model costs 40% of a rival’s, and 5% of another’s, buyers gain leverage in every renewal. Even if the claims shrink under independent testing, the reference point now exists. Sales teams at OpenAI and Anthropic will hear the 5% figure in negotiations, whether or not it survives scrutiny.
The second effect is architectural. Licensed weights plus a router like Decision 1.1 let an enterprise assemble a mixed fleet of models. A bank could run a customized coding model, a transcription model and a classifier, all inside Azure, and call outside models only for the hardest tasks. That shifts spending from per-token API bills to licenses and compute inside one cloud.
The third effect is political. Microsoft’s AI relationships already draw attention in Washington, as our coverage of a House panel pressing Microsoft over a reported $500M ByteDance deal shows. Owning more of its own model stack gives Microsoft a story about control, and it reduces exposure to a single outside supplier. It also puts more scrutiny on the training data, safety testing and disclosure behind those models.
The sources contain no stock reaction, revenue forecast or customer count tied to the launch. Any claim about market share shifts would be speculation at this stage. What the record does show is a vendor that now sells models, a cloud and an assistant, and that bundle is hard for a pure model lab to match.
Historical Context: From Reseller to Model Maker #
India Today notes that Microsoft has long relied on OpenAI to power its AI tools. ChatGPT launched in November 2022 and turned that partnership into the center of Microsoft’s AI pitch, as Ticker News recalls. For years, Microsoft’s Copilot products and Azure AI services drew their headline capability from models built elsewhere.
The MAI family changes the framing. Ticker News describes the launch as a step to expand Microsoft’s technology beyond the OpenAI partnership, and Suleyman outlined a goal of building proprietary capabilities and reducing reliance on outside partners. Enterprise AM used similar language, saying Microsoft is trying to rely less on OpenAI.
Seen over a longer arc, this is a familiar pattern. Large platform companies often start by distributing a partner’s breakthrough, then build in-house versions once the market proves out. Whether its in-house models match the original is the open question, and no source reviewed here says the partnership has ended.
What Is Confirmed and What Is Not #
With this many claims in flight, a ledger helps. The table separates what Microsoft’s materials and named outlets support from what still lacks evidence.
| Claim | Status | Basis |
|---|---|---|
| Microsoft announced new in-house models at Ignite on Nov. 17 | Confirmed | Ticker News, Axios, India Today |
| Exact six-model lineup | Unclear | Sources name seven or eight products |
| MAI Cyber 2 Flash has 12 billion active parameters | Stated by Microsoft | Microsoft announcement |
| MAI Code 2 reaches 98% of GPT-6.1 Sol at 40% of cost | Vendor claim | Microsoft keynote, no independent test |
| MAI Code 2 reaches 96% of Claude Opus 5.5 at 5% of cost | Vendor claim | Reported by India Today |
| Customers can license weights and run them in Azure | Confirmed | Microsoft and press reports |
| Customer pricing for any model | Not announced | Ticker News says details are not released |
Two gaps stand out. First, nobody has published prices, so every cost comparison compares a known rival price to an unknown Microsoft one. Second, no outside lab has verified the benchmark averages. Until both gaps close, the launch is best read as a strong opening bid, not a settled result.
Five Predictions for the Next 12 Months #
These forecasts are analysis, not reported facts, and they rest on the pattern of the launch so far.
- Independent benchmarks will trim the headline claims. Vendor averages rarely survive outside testing intact. Expect the 98% and 96% figures to move, and expect Microsoft to argue the case with real-world task data.
- OpenAI and Anthropic will answer on price within weeks. OpenAI already cut Astra’s price, according to India Today. More cuts, bundles and discounts for committed spend look likely.
- Routing will become a standard enterprise layer. Decision 1.1 points to a market where companies buy a router first and pick models second.
- Licensed weights will spread to more models. Microsoft launched the licensing program, and customer demand for control suggests it will widen the list.
- The general-purpose frontier model will arrive. Suleyman confirmed work on one, per Ticker News. When it lands, the narrow-model strategy will meet its real test against the biggest rivals.
The takeaway: Microsoft has stopped being only a distributor of other labs’ models. Whether the MAI family wins on merit depends on prices Microsoft has not published and tests nobody outside the company has run.
FAQ: Microsoft’s New MAI Models #
How many AI models did Microsoft announce at Ignite?
Ticker News and India Today say six. The named products number more than that, including MAI Code 2, MAI Code 2 Flash, MAI Image 2.7, MAI Image 2.7 Flash, MAI Cyber 2 Flash, MAI Realtime 1, MAI Transcribe 2 and Microsoft-Decision-1.1. Microsoft’s own page lists a slightly different set, so the exact six cannot be fixed from the available record.
When and where did Microsoft announce them?
Microsoft unveiled the models at its Ignite conference in San Francisco on November 17, 2026.
What is MAI Code 2 built for?
Microsoft designed MAI Code 2 for complex, long-horizon, multi-file coding tasks, and says it can reason across an entire repository. Reports say it will come to GitHub Copilot.
Is MAI Code 2 really 98% as good as GPT-6.1 Sol at 40% of the cost?
That is Microsoft’s claim, stated by Mustafa Suleyman in the keynote. It averages four benchmarks for everyday enterprise coding. The available record shows no independent confirmation, so treat it as a vendor result until outside testers publish their own numbers.
Can companies customize the models with their own data?
Yes. Enterprise customers can license Microsoft’s model weights and customize the models with proprietary data. They must deploy the models within Azure.
What do the new models cost?
Microsoft has not published customer prices in the sources reviewed, and Ticker News reports that pricing details are not yet released. Only relative claims, such as 40% of a rival’s cost, are public.
What does MAI Cyber 2 Flash offer defenders?
Microsoft says it has 12 billion active parameters. India Today reports Microsoft’s claim that it beats GPT-6 Astra on a major security test at under 2% of the per-task cost, though that result has no independent check yet.
Does this end Microsoft’s reliance on OpenAI?
No source reviewed says so. Press coverage frames the launch as a step toward proprietary capability and less reliance on OpenAI, not as an end to the partnership.