⚡ Quick verdict
- What happened → At Ignite 2026 in San Francisco on 17 November 2026, Microsoft AI launched seven of its own MAI models — Code 2, Code 2 Flash, Cyber 2 Flash, Image 2.7, Realtime 1, Transcribe 2 and Decision 1.1 — plus Copilot Workspaces, GPT-6.2 on Foundry, the Maia 300 chip and an AI Transparency Initiative.
- The pitch → "Frontier performance. A fraction of today's prices." Every headline number is a price-performance claim: Code 2 at 98% of GPT-6.1 Sol's coding quality for about 40% of the price, Cyber 2 Flash at under 2% of GPT-6 Astra's cost per task, Image 2.7 ranked #1 for image editing at about a fifth of GPT Image 2.6's price.
- The catch → Almost every headline figure is Microsoft's own, measured on benchmarks Microsoft chose — and the Code 2 price comparison uses a 50% promotional price, so at full price it is closer to 80% of Sol's cost. No independent benchmark of the coding or security models existed when we wrote this on 18 November 2026.
- What's actually usable today → MAI Code 2 and Code 2 Flash are generally available in GitHub Copilot right now, and Decision 1.1 is generally available. The rest is preview, limited access, or gated behind Microsoft's Frontier program. Test Code 2 on your own code before believing any number in this article.
- Best for → Developers who live in GitHub Copilot, enterprises running agents around the clock, security teams already inside Microsoft's stack, and anyone whose AI bill has become a line item worth negotiating. Skip the hype and judge the models on your own workload.
We may earn a commission if you buy through our links, at no extra cost to you.
Microsoft opened Ignite 2026 in San Francisco on 17 November 2026 with a message that was less about a single product and more about a strategy: the company that spent years selling you other people's models — OpenAI's GPT series, Anthropic's Claude, Meta's Llama — now wants to sell you its own. Satya Nadella's opening keynote, and a separate keynote from Microsoft AI chief Mustafa Suleyman, introduced seven new models from Microsoft AI (MAI). The pitch, in Microsoft's own words, was "frontier performance at a fraction of today's prices."
That claim is aimed at a real problem. AI agents that run around the clock — coding, triaging security alerts, answering customers, transcribing calls — burn tokens at a rate that makes frontier-model pricing painful. A model that is 90-something percent as good for a tenth of the price changes what a business can afford to automate. Microsoft's bet is that most enterprise work does not need the smartest model in the world — it needs a good-enough model that is cheap, fast, and tuned for one job.
This is a deep-dive review and analysis, not a repeat of the news. Our same-day news item — Microsoft unveils 7 MAI models at Ignite 2026 — covers the announcement itself. Here we go further: what each model actually does, where you can use it today, which numbers are solid and which are marketing, what is genuinely new, and who should care. Everything here is based on Microsoft's own announcements of 17 November 2026, the Microsoft AI keynote transcript, and same-day reporting from Reuters, Axios, The Verge and others — checked on 18 November 2026. Where a figure is Microsoft's own claim, we say so.
The seven MAI models at a glance
Microsoft AI describes the launch as "a family of seven new world class models." They cover six kinds of work: coding, cybersecurity, image generation and editing, live voice conversation, transcription, and model routing. Two of them have a cheaper "Flash" sibling. The table below is the whole announcement in one place, with availability as Microsoft stated it on 17 November 2026.
The seven MAI models — Ignite 2026
- MAI Code 2
- Everyday enterprise coding — deep debugging, multi-file changes, whole-repo reasoning. Microsoft claims 98% of GPT-6.1 Sol quality at ~40% of the price (promotional pricing — see the caveats). GA in GitHub Copilot.
- MAI Code 2 Flash
- High-volume, repetitive coding — explanations, targeted edits, routine fixes. Microsoft claims performance comparable to GPT-6 Luna at 65% of the price and Claude Haiku 5.5 at 20%. GA in Copilot, including Free and Student; limited-access preview in Foundry.
- MAI Cyber 2 Flash
- Finding real vulnerabilities in code and helping defenders investigate attacks. Microsoft claims 92.3% on CyberGym-E2E vs 91.5% for GPT-6 Astra, at under 2% of the cost per task. Verified defenders only.
- MAI Image 2.7 (+ Flash)
- Image editing and generation. Ranked #1 on Artificial Analysis image editing. About $35 per 1,000 images in public preview on Foundry; Flash about $10.
- MAI Realtime 1
- Microsoft's first live speech-to-speech model for natural, low-latency voice conversation. From $0.01 per audio minute. Demo via MAI Labs; limited-access preview in Foundry.
- MAI Transcribe 2
- Multilingual transcription, powering Dragon Copilot Transcribe in healthcare. Microsoft says it handles about 10 million patient encounters a month.
- Microsoft Decision 1.1
- Routing model that sends each request to the cheapest model that can handle it. Now GA, used by more than 4,000 customers, on Foundry, OpenRouter and Vercel.
Three patterns run through the whole family. First, every model is specialised: one job, tuned hard, instead of one giant model that does everything. Second, every model is priced to undercut a named frontier rival. Third, the family is built to be customised: Microsoft announced MAI Licensed Weights, which lets trusted enterprise customers get access to the model weights through its Frontier Co. programme, so they can tune the models on their own data inside their own compliance boundary.
The rest of this review takes the family one model at a time, then steps back to the bigger Ignite story: Copilot Workspaces, Foundry's new frontier models, the Maia chips, and the governance push. We finish with the honest verdict — what is real, what is marketing, and what to do on Monday morning.
MAI Code 2: the model that matters most
If you remember one model from this launch, make it MAI Code 2 — because it is the one you can actually use today, and the one Microsoft put at the centre of the keynote. Code 2 is built for the coding work enterprises do every day: taking a project to production, fixing deep bugs, making changes across many files, reasoning over a whole repository, and polishing user interfaces. It is not aimed at frontier research or very long-horizon agent work; Microsoft is explicit that the hardest problems are not its target. It is aimed at the 90% of coding that is routine, repetitive, and currently billed at frontier-model prices.
The headline claim is the one you will see everywhere today: averaged across four coding benchmarks — DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1 — Microsoft says Code 2 delivers 98% of GPT-6.1 Sol's performance at about 40% of the cost, and 96% of Claude Opus 5.5's performance at about 5% of the cost. Those benchmarks were chosen to mirror the most frequent tasks in GitHub Copilot. An average across four tests can hide a test where the model trails badly, and all four tests were selected by the vendor making the claim — that is the first caveat, and it applies to every number in this review.
The second caveat is bigger, and it is the one that changes the arithmetic. A footnote in Microsoft's own Code 2 post says the model's cost "is shown at its 50% promotional price." In other words, the 40% figure is the promotional price, not the list price. Independent analysis by FourWeekMBA, working from Microsoft's own footnote, puts the full-price comparison at roughly 80% of GPT-6.1 Sol's cost and about 10% of Claude Opus 5.5's — still cheaper, but a very different headline. The Opus comparison stays dramatic even at full price, because Opus 5.5 is one of the most expensive models on the market. The Sol comparison, at full price, is a modest discount rather than a revolution. Microsoft has not published the list price for Code 2 or how long the promotion runs, so check the official GitHub Copilot and Foundry pricing pages before you budget around the 40% figure.
None of that makes Code 2 a bad model — it makes it a model whose price story needs reading in full. If Code 2 really does 98% of Sol's work on everyday enterprise coding, that is a genuinely useful result even at 80% of the price, because it runs inside the tool developers already use. And there is a real-world data point in Microsoft's favour: AT&T's custom version of Code 2 Flash, tuned for the company's own work, reportedly outperforms GPT-6.1 Sol at a standard 5G network analysis while being about 50 times cheaper. A model tuned for your codebase, on your tasks, can beat a general model that has never seen them.
Availability is the strongest part of the Code 2 story. Both coding models are generally available in GitHub Copilot from 17 November 2026, across VS Code, Visual Studio, the Copilot CLI, the cloud agent, GitHub.com and the mobile apps. Code 2 is on Copilot Pro, Pro+, Max, Business and Enterprise plans. Code 2 Flash is on those plans too, plus Copilot Free and Student — which means a student or a hobbyist on the free tier can try a Microsoft-built coding model today at no cost. That is a smart distribution move: the model that needs the most testing gets the widest possible audience. If you want to know whether Code 2 is any good, you do not need to take Microsoft's word for it. Open Copilot, pick Code 2 in the model picker, and run it on your own hardest bug. Your own repository is a better benchmark than any leaderboard.
For the wider market, Code 2 challenges the idea that the best coding model must come from OpenAI or Anthropic — our best AI coding assistants roundup compares the purpose-built tools, and until this week every serious entry in it was built on someone else's model. Microsoft is now a model builder in its own right, and Copilot is the delivery vehicle.
MAI Code 2 Flash: the workhorse sibling
Code 2 Flash is the lighter, faster, cheaper sibling, and for many teams it may turn out to be the more important model. It targets the frequent, repetitive jobs: explaining code, applying small targeted edits, fixing routine problems, and iterating quickly. Microsoft compares it with GPT-6 Luna at 65% of that model's price and Claude Haiku 5.5 at 20% of the price, across three coding benchmarks — and the same promotional-pricing footnote applies, so at full price those figures would be roughly 130% of Luna's cost and 40% of Haiku 5.5's, by the same independent arithmetic. Read that carefully: at list price, Code 2 Flash may actually cost more than GPT-6 Luna for the same work. The promotional price is doing real work in Microsoft's headline, and anyone building a budget on it should check the official pricing first.
What makes Flash interesting is the tuning story. Because it is a smaller, specialised model, it is cheap enough to customise — and Microsoft's AT&T example is the proof of concept: a custom version of Flash, tuned for one company's network analysis, beating a frontier model at 50 times lower cost. Combined with MAI Licensed Weights and the Frontier Tuning programme, Flash is the entry point to a coding model that knows your codebase, your conventions, and your infrastructure. That is a bigger idea than a price cut, and it is the one worth watching over the next year.
Flash is also the model with the widest reach: it is in GitHub Copilot on every plan including Free and Student, and in limited-access preview in Microsoft Foundry for developers building their own tools. If you are a student, a solo developer, or a small team, Flash is the free way to test whether Microsoft's models are worth your time. Try it on the boring work — the explanations, the small fixes, the repetitive edits — and see whether it is fast and accurate enough to keep. That is the job it was built for, and it is the job where a cheaper model earns its place.
MAI Cyber 2 Flash: security at 2% of the cost
The most eye-catching number in the whole launch belongs to the security model. Microsoft says MAI Cyber 2 Flash scores 92.3% on CyberGym-E2E — a benchmark for finding real vulnerabilities in code and proving they are real — against 91.5% for GPT-6 Astra, at under 2% of the cost per task. Unlike the Code 2 comparison, this one is priced at list: $0.10 per million input tokens and $0.50 per million output tokens, against $10 and $50 for GPT-6 Astra. That is a printed price, not a promotional footnote — the cleanest cost claim in the announcement. Security work is the perfect use case for a cheap specialised model, because it is high-volume, always-on, and mostly routine.
The catch is access. Cyber 2 Flash is not a model you can download or call from an API today. Microsoft says it is "coming exclusively to Microsoft security products," and the only date it has set is mid-December 2026, when it replaces the existing model in the MAI profile of MDASH — Microsoft's security data platform — for eligible customers. Only verified defenders can get access, after registering their interest. That is a deliberate choice: a model that finds vulnerabilities is also a model that could be misused to find them for the wrong reasons, and Microsoft is keeping it inside its own security stack. For most readers, that means Cyber 2 Flash is a model to watch, not a model to try. If your organisation already runs Microsoft's security products, ask your Microsoft contact when it reaches your tenant. If it does not, this model will not change your week.
Cyber 2 Flash succeeds MAI Cyber 1.1 Flash from October 2026, so the security line is iterating fast. Its 1-million-token context window matters — investigating a vulnerability means reading a lot of code, logs and history. But the benchmark is Microsoft's own chosen test, with no independent evaluation yet: the price is printed and impressive, the score is a vendor claim, and the real test will be whether security teams notice a difference when it ships in December.
MAI Image 2.7: the one independent win
MAI Image 2.7 is the model with the strongest independent evidence in the whole launch. Artificial Analysis ranks it first on its AA-Image-Editing v2.1 leaderboard, with 1,215 Elo points built from 5,478 head-to-head votes, ahead of GPT Image 2.6 Sunburst at its maximum setting. Microsoft says the model beats GPT Image 2.6 at about one fifth of the price — and Artificial Analysis lists the model's own price at $34.82 per 1,000 images, with the faster Flash variant at $10.24 per 1,000. That ranking is not Microsoft's own test; it is the only headline in this launch with independent backing. Our earlier coverage of the Image 2.7 preview has more background.
Image 2.7 is built for targeted editing: changing one part of an image while keeping the rest of the scene intact — fix the background, change the outfit, swap the product colour, keep the face. Microsoft says the model is designed for exactly that. It is also now the default model behind Bing Image Creator, which you can try free through Bing Image Creator.
Both image models are in public preview on Microsoft Foundry, which makes them the most accessible new models in the family after the coding pair. The pricing is printed and straightforward: about $35 per 1,000 images for Image 2.7, about $10 for Image 2.7 Flash. At those prices, the Flash variant is the interesting one for high-volume work — social media images, product shots, marketing variations — where you need many images fast and the last 5% of quality does not matter. The full model is for the edits where quality does matter. One caution: a #1 ranking on an image-editing leaderboard measures one kind of task, judged by head-to-head votes. It does not prove the model is best at text-to-image generation, photorealism, or brand-consistent product shots. Check the Artificial Analysis leaderboard for the current ranking, and test the model on your own images before you switch a workflow to it. For the wider field, our best AI image generators roundup compares the alternatives.
MAI Realtime 1: Microsoft's first speech-to-speech model
MAI Realtime 1 is Microsoft's first live speech-to-speech model — it listens and speaks in real time, with low latency, rather than transcribing speech to text, thinking, and then reading an answer aloud. A speech-to-speech model can interrupt, pause, change tone, and respond to the way you say something, not just the words. It is the technology behind natural voice assistants and customer-service bots that sound like a person rather than a voicemail menu. Microsoft says Realtime 1 beats OpenAI's GPT Realtime 2.1 on quality, speed and price, in public benchmarks and its own product evaluations — partly Microsoft's own testing, so treat the "beats" as a vendor claim until independent voice benchmarks weigh in.
The price is the clearest part of the story: Realtime 1 starts at $0.01 per audio minute, with reasoning billed separately at $1 per million input tokens and $5 per million output tokens. One cent a minute is a number a business can do arithmetic with. A thousand minutes of voice conversation costs about $10, before reasoning. That is the kind of pricing that makes voice agents affordable for customer service, appointment booking, and phone support — the jobs where a human costs dollars per minute and a frontier voice model costs enough to make the business case marginal. Microsoft is betting that a cent a minute is the price point where voice agents stop being a demo and start being a deployment.
Access, however, is limited. Realtime 1 is offered through MAI Labs, with a demo and an interest form, and in a limited-access preview in Foundry — and Microsoft says preview customers will not be billed until a public preview begins. That is a soft launch, not a general release. Voice is the newest, least proven part of the family, and the quality bar is highest here, because a voice agent that misunderstands a customer is worse than no agent at all. Judge it by whether it handles interruptions, accents, and background noise. For the current state of the field, our best AI chatbot roundup covers the voice-capable assistants you can actually use today.
MAI Transcribe 2: the quiet workhorse in healthcare
MAI Transcribe 2 is the least flashy model in the family and possibly the most widely used already — because it is already working. It powers Dragon Copilot Transcribe, the transcription engine inside Microsoft's Dragon Copilot healthcare product, and Microsoft says it already handles about 10 million patient encounters a month. Microsoft calls it "the world's fastest and most accurate multilingual transcription model" — a superlative that should be read as marketing, since no independent test has crowned it — but the usage number is a real signal: a model transcribing 10 million clinical conversations a month is infrastructure, not a demo. Microsoft also says the model cuts multilingual error rates by about half, and that roughly a third of US physicians already use Dragon Copilot — both Microsoft's own claims, not independent findings.
What Transcribe 2 does is turn speech into accurate text in many languages, fast enough to keep up with a live conversation — and in a clinical setting, a transcription error is a patient-safety issue, not an embarrassment, so the "most accurate" claim needs independent verification more than most. For anyone outside healthcare, Transcribe 2 is not a product you will use directly — you will meet it inside Dragon Copilot, or not at all. Try it through our Dragon Copilot link if you work in healthcare.
Microsoft Decision 1.1: the router that makes the rest cheap
Decision 1.1 is the least glamorous model in the family and the one that makes the whole strategy work. It is a routing model: it looks at each incoming request and decides which model should handle it — the cheap one, or the expensive one. Send the easy questions to a small model, the hard ones to a frontier model, and you get most of the quality for a fraction of the cost. Microsoft says Decision 1.1 is 50 times faster and 79 times cheaper than GPT-6.1 Sol for this kind of micro-decision work — Microsoft's own comparison, so read it as a claim — and it is priced at $0.042 per million input tokens. Artificial Analysis scores it 68 on its Decision Index, up from 62 for the previous version.
Decision 1.1 is the only model in the family besides the coding pair that is generally available today, and it has the widest distribution: it is on Foundry, on OpenRouter, and on Vercel, and Microsoft says more than 4,000 customers already use it. If you run an AI product that calls models thousands of times a day, Decision 1.1 is the model to test first, because it can cut your bill without changing what your users see. You do not have to choose between Microsoft's models and the frontier models — you use both, and a router decides which one each request deserves.
Copilot Workspaces: Copilot becomes the OS for work
The models were only half the keynote. The other half was Copilot itself — Nadella's framing was deliberately grand: "Copilot is the new OS for work." The rebuilt Copilot is organised around four modes — Chat, Cowork, Code and Autopilot — in one experience, with Work IQ underneath and a new Microsoft IQ platform bringing together data from Microsoft 365, Fabric, Foundry and third-party apps, with about 350 connectors. The idea is that Copilot stops being an assistant you ask questions and becomes the place where work happens.
Copilot Workspaces is the centrepiece of that idea. Until now, Microsoft says, Copilot has mostly been a single-player world: you and your agent loops. Workspaces make it multiplayer — people, their agents, and other people, all driving towards a common goal with shared context. A workspace is a shared space where a team and its agents work on the same project, with the same information, and the workspace carries on in Teams. The old Copilot Notebooks become Workspaces in this new model. Alongside them, Autopilot handles long-running, repetitive tasks and projects without step-by-step instructions — you give it a goal and it keeps working — and Copilot Code lets anyone build dashboards, trackers and applications in natural language, with enterprise security and governance applied by default.
The honest version, and the one that matters for planning, is the availability. Workspaces are rolling out through Microsoft's Frontier programme — a gated preview, not general availability — and no pricing has been published. Copilot Code is expected to reach general availability "in the coming weeks," which is not a date. Autopilot is Frontier-gated too. Independent coverage of the keynote puts it plainly: about half of the Ignite announcements are usable today, and about half are a Frontier preview with no date. Do not plan a rollout around Workspaces or Autopilot yet. If your organisation is in the Frontier programme, try a workspace on a real project and see whether the shared context actually helps a team, or whether it just adds another place to check. If you are not in the programme, the practical near-term work is the models — Code 2 in Copilot, Decision 1.1 in your stack — not the workspaces. Try Copilot through our Microsoft Copilot link, and check the official Microsoft 365 pages for the current Workspace rollout status.
Microsoft is also extending Copilot for specific roles — sales, legal, finance, service — including a partnership to bring Salesforce into Copilot. For sales teams, our best AI SDR tools comparison covers the purpose-built tools that do that job today, and our best AI SEO tools roundup does the same for marketing work.
Foundry: GPT-6.2, Meta's models, and Microsoft's own weights
One of the most important things Microsoft said at Ignite is also the easiest to misread: Microsoft is not replacing the frontier models. It is selling them alongside its own. Nadella announced that GPT-6.2 — OpenAI's newest model — is coming to Foundry, and that Meta's latest models are coming to Foundry as well. Foundry remains the place where Microsoft sells you everyone else's models alongside its own, and in Copilot, users can pick OpenAI or Anthropic models, subject to their organisation's policy. That is the honest shape of the strategy: MAI models for high-volume routine work, frontier models where quality matters most, and a router deciding which is which.
That framing is worth keeping in mind when you read the price claims: the company that sells both sides of the comparison is also the company that chose the tests. The counterweight is that the strategy only works if the MAI models are actually good — a cheap model that produces bad code is not a bargain at any price, and Microsoft knows it, which is why the models are in GitHub Copilot, where millions of developers will test them on real work within days.
The other Foundry announcement with long-term significance is MAI Licensed Weights. Trusted customers will now be able to get access to Microsoft AI model weights through the Frontier Co. programme — meaning they can take the weights, tune them on their own data, and run them inside their own compliance boundary. Most providers rent you API access; giving customers the weights is a different kind of trust, and a different kind of lock-in. The programme is gated to trusted customers, with no public date for wider access, so treat it as a direction of travel rather than a product you can buy today. But for regulated industries — healthcare, finance, government — where data cannot leave the building, licensed weights may turn out to be the most important announcement of the keynote. Check the official Foundry pages for the current status before planning around it.
Maia 200 in production, Maia 300 announced
Behind the models sits the silicon. Microsoft says each of the new MAI models is now running in production on its second-generation Maia 200 AI accelerator for datacentres — the company's own chip, built to run its own models efficiently. That is the infrastructure story behind the price story: Microsoft can undercut frontier-model pricing partly because it controls the hardware the models run on. The next chip, Maia 300, was announced at Ignite with a claim of 40% more tokens per watt than leading GPU-based systems — Microsoft's own claim, about a chip that is not yet shipping, so treat it as a roadmap number rather than a measured result.
Microsoft also announced general availability of its next-generation Cobalt 200 virtual machines, with expansion to 23 Azure regions planned by the end of 2026. The chip and VM announcements are part of the same argument: Microsoft is building the whole stack, from silicon to model to workspace, and the price advantage comes from owning every layer. That is a structural advantage, not a one-off promotion — and it is the reason the MAI pricing strategy is credible even where the headline percentages are not.
The AI Transparency Initiative: governance gets a real answer
The third strand of the keynote was governance, and it is the one that will matter most to large organisations. Microsoft announced the AI Transparency Initiative, a programme with more than 40 participating organisations — including Adobe, Mastercard, Accenture, KPMG and Volvo Cars — to develop common, open, industry-wide standards for AI transparency, security and governance, including methods for independently verifying what AI systems do. The initiative has three parts: expanding Microsoft's collaboration in NIST's Agent Standards Initiative, developing the standards with the participating organisations, and incorporating them into Microsoft's own products, including MAI models, Foundry and Copilot.
This is a bet that the next competitive advantage in AI will not be raw capability but trust. As agents start taking actions — spending money, sending emails, changing records — organisations need to verify what the agent did, why it did it, and whether it was allowed to. The initiative is new, the standards do not exist yet, and 40 organisations signing up is a beginning, not a result. Alongside it, Microsoft announced Agent 365, its governance layer for agents — Microsoft says more than 50 million agents already run inside it — with cost management that tracks consumption by user, service, agent and model, with limits, policies and alerts.
What is genuinely new — and what is marketing
After a launch this size, the useful question is not "was it impressive" but "what actually changed." Here is our honest split, one day after the announcement.
Genuinely new. Microsoft is now a first-party model builder at scale, with seven specialised models across six kinds of work. Two of them are generally available in GitHub Copilot — a product you can try today, not a slide deck. The image model has a genuinely independent #1 ranking, the routing model is GA on three platforms, licensed weights are a real shift in how models are sold, and the Maia 200 is running the models in production on Microsoft's own silicon. Those are facts, not claims.
Mostly marketing. The headline price percentages are the weakest part of the launch. The Code 2 comparison uses promotional pricing, the benchmarks were chosen by the vendor, the "beats GPT" claims are partly Microsoft's own evaluations, and the superlatives — "world's fastest and most accurate," "frontier performance" — are advertising copy, not measurements. Workspaces, Autopilot and Copilot Code are previews with no dates and no prices, the Maia 300's 40% figure is a claim about unshipped hardware, and the Transparency Initiative is a promise, not a standard. That is why this review keeps saying "Microsoft says" instead of stating the numbers as facts.
In between. The specialisation argument is the most credible part of Microsoft's pitch — the AT&T example suggests a tuned small model can beat a frontier model on a narrow task. But a tuned model only wins on the task it was tuned for. The value of the MAI family will be decided by how well the tuning tools work on your own data, outside Microsoft's chosen tests. That is a question for the next six months, not for launch day.
What is still unproven
Four things need to happen before anyone can give the MAI family a final verdict. First, independent benchmarks — no independent evaluation of Code 2, Cyber 2 Flash or Realtime 1 had been published when we wrote this on 18 November 2026, and the first serious tests by someone who does not sell either side will tell us more than any keynote. Second, the pricing: the promotional footnote on Code 2 means the real price is unknown, and the prices for Workspaces, Autopilot and licensed weights have not been published at all. Third, the previews: most of the family is in preview or limited access, and preview performance is not production performance. Fourth, the customisation story: if tuning a small model on your own data turns out to be easy and effective, the MAI family becomes a platform; if it is hard and expensive, it is just a set of cheaper models.
One more thing to watch, and it is not in the keynote: OpenAI and Anthropic will respond. GPT-6.2 is coming to Foundry, but it is also a competitor's product, and its pricing will set the bar Microsoft has to beat. The numbers in this review are November 2026 numbers — check the official pricing pages before you decide, because they will have changed by the time you read this.
Who each model is for
Who should use which MAI model — November 2026
- MAI Code 2
- Developers in GitHub Copilot doing everyday enterprise coding. Try it on your own repo this week; it is GA today.
- MAI Code 2 Flash
- Students, solo developers and teams on Copilot Free or Student, and anyone with high-volume repetitive coding work.
- MAI Cyber 2 Flash
- Security teams already inside Microsoft's security products. Verified defenders only, arriving from mid-December 2026.
- MAI Image 2.7
- Marketers, designers and content teams who need precise image edits. Public preview on Foundry; Flash for high-volume work.
- MAI Realtime 1
- Teams building voice agents for customer service or phone support, once the preview opens. Join the MAI Labs waitlist.
- MAI Transcribe 2
- Healthcare organisations using Dragon Copilot. Already live, per Microsoft.
- Decision 1.1
- Any team running AI at scale that wants to cut its model bill. GA on Foundry, OpenRouter and Vercel.
MAI models vs the field
Here is how the family sits against the alternatives most US readers actually use — with the caveat that the MAI figures are Microsoft's own claims. Our best AI chatbot roundup and best AI coding assistants roundup have the full comparisons.
MAI family vs the alternatives — November 2026
- Coding
- MAI Code 2 (GA in Copilot) vs GPT-6.1 Sol and Claude Opus 5.5 — Microsoft claims 98% and 96% of their quality, but the price comparison is promotional and no independent test exists yet.
- Image editing
- MAI Image 2.7 is ranked #1 on Artificial Analysis image editing, ahead of GPT Image 2.6 Sunburst — the strongest independent result in the launch.
- Voice
- MAI Realtime 1 vs OpenAI GPT Realtime 2.1 — Microsoft claims better quality, speed and price from $0.01 per audio minute, but it is still in limited preview.
- Routing
- Decision 1.1 vs running every request on a frontier model — the most cost-effective pattern in the family, and it is GA today.
- General chat
- Microsoft is not competing here — Copilot lets users pick OpenAI or Anthropic models, and GPT-6.2 is coming to Foundry. For everyday chat, the established assistants remain the better choice.
Microsoft MAI Models
Strengths
- Code 2 and Code 2 Flash are generally available in GitHub Copilot today — not a slide deck, a product you can try.
- Image 2.7 has a genuinely independent #1 ranking on Artificial Analysis image editing.
- Printed prices for image, voice and security models make the cost case concrete, not hypothetical.
- Decision 1.1 is GA on three platforms and is the easiest cost saving in the family.
- Specialisation plus licensed weights is a real strategy, not just a price cut — tuned small models can beat frontier models on narrow tasks.
- Full-stack ownership, from Maia silicon to Copilot, makes the pricing credible over time.
Weaknesses
- Almost every headline number is Microsoft's own, on benchmarks Microsoft chose — no independent test of the coding or security models yet.
- The Code 2 price claim uses a 50% promotional price; at full price the Sol comparison is roughly 80%, not 40%.
- Most of the family is preview, limited access, or Frontier-gated — Workspaces, Realtime 1, Cyber 2 Flash and Copilot Code are not GA.
- No pricing published for Workspaces, Autopilot or licensed weights.
- Superlatives like "world's fastest and most accurate" are marketing, not measurements.
- Customisation is the big promise and the least tested part — tuning may be harder than the keynote suggests.
What to watch before you trust it
Four things will decide whether this launch is remembered as a turning point or a keynote. First, the independent benchmarks: the first serious tests of Code 2 against GPT-6.1 Sol and Claude Opus 5.5, run by someone who does not sell either side, on real repositories rather than chosen tests. Second, the real pricing: the list price for Code 2 once the promotion ends, and the prices for Workspaces, Autopilot and licensed weights when published. Third, the previews: whether Realtime 1, Cyber 2 Flash and Workspaces reach general availability and work as well in production as in a keynote demo. Fourth, the tuning story: whether ordinary enterprises can customise these models on their own data and get the AT&T-style results Microsoft showcased.
Watch the competitors too: Anthropic, Google and Meta will all respond to a price war, and the most likely outcome is that everyone gets cheaper. For the tools that already do these jobs well, our coding assistants, image generators and chatbot roundups are the place to compare, and our best AI resume builders roundup covers a different kind of everyday AI work. We will update this review as the independent tests come in.
Our verdict
Microsoft's Ignite 2026 launch is the most important AI announcement of the month, and the hardest to score on day one. The models are real, the availability is real for the coding pair and the router, and the strategy is coherent: specialised models, cheap enough to run all day, tuned for your work, routed by a model that knows when to spend and when to save. The image model's independent #1 ranking is a genuine achievement. If you are a developer in GitHub Copilot, try Code 2 this week on your own code. If you run AI at scale, test Decision 1.1 against your current bill. If you make images, try Image 2.7 on your own edits.
But the headline numbers are not the story. They are Microsoft's own claims, measured on Microsoft's chosen tests, with a promotional price doing real work in the most-quoted figure. The rest of the family is preview, gated, or undated. And the superlatives should be ignored until independent testing exists. That is not a dismissal — it is the same standard we apply to every launch-day product, including the ones we like. Microsoft has earned a serious hearing with this launch. It has not yet earned the benefit of the doubt on the numbers. The models that survive the next six months of independent testing will be the ones worth building on — and on the evidence of day one, Code 2 and Image 2.7 are the two most likely to survive.
Bottom line: a real model family with a real strategy and a real price advantage — but read the footnotes, test the models on your own work, and wait for independent benchmarks before you believe the percentages. The best move this week is free: open Copilot, pick Code 2, and run it on your hardest bug.
Frequently asked questions
What are Microsoft's seven MAI models announced at Ignite 2026?
The seven models announced on 17 November 2026 are MAI Code 2 and MAI Code 2 Flash for coding, MAI Cyber 2 Flash for cybersecurity, MAI Image 2.7 (plus a 2.7 Flash variant) for image editing and generation, MAI Realtime 1 for live speech-to-speech conversation, MAI Transcribe 2, which powers Dragon Copilot Transcribe for healthcare, and Microsoft Decision 1.1, a routing model that is now generally available. They are Microsoft AI's own in-house models, built to be cheaper to run than the frontier models from OpenAI, Anthropic and Meta that Microsoft also sells through Foundry.
Is MAI Code 2 really 98% as good as GPT-6.1 Sol at 40% of the price?
That is Microsoft's own claim, not an independent result. Microsoft says MAI Code 2 averages 98% of GPT-6.1 Sol's quality at about 40% of the price across four coding benchmarks it chose — DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1 — and 96% of Claude Opus 5.5 at about 5% of the price. The important footnote is that Microsoft's own pricing uses a 50% promotional price for Code 2, so at full list price the Sol comparison would be closer to 80% of the cost, by independent arithmetic. No independent benchmark of Code 2 had been published when we wrote this review on 18 November 2026, so treat the headline as a vendor claim to be tested.
How can I try MAI Code 2 today?
MAI Code 2 and MAI Code 2 Flash are generally available in GitHub Copilot from 17 November 2026 — Code 2 on Copilot Pro, Pro+, Max, Business and Enterprise plans, and Code 2 Flash on those plans plus Copilot Free and Student. They work in VS Code, Visual Studio, the Copilot CLI, the cloud agent, GitHub.com and the mobile apps. Code 2 Flash is also in limited-access preview in Microsoft Foundry. Pick the model in the Copilot model picker and check the official GitHub Copilot pages for current availability and pricing before you commit to a plan.
What is Copilot Workspaces?
Copilot Workspaces is Microsoft's new shared workspace inside Copilot where people and AI agents work on the same project with the same context — a multiplayer version of Copilot, replacing the old single-player Copilot Notebooks idea. A workspace carries on in Teams, and long-running Autopilot tasks can keep working on a project without step-by-step instructions. It is rolling out through Microsoft's Frontier program, not generally available yet, and no pricing has been published — so plan for a preview, not a rollout, and check the official Microsoft 365 pages for the current status.
Is MAI Image 2.7 really the best image editing model?
On one independent leaderboard, yes — Artificial Analysis ranks MAI Image 2.7 first on its AA-Image-Editing v2.1 leaderboard with 1,215 Elo points, ahead of GPT Image 2.6 Sunburst at its maximum setting. Microsoft prices it at about $35 per 1,000 images in public preview on Foundry, with the faster Image 2.7 Flash at about $10 per 1,000. That ranking is a real independent data point, but it measures one kind of task — image editing judged by head-to-head votes — so it does not prove the model is best at every image job. Check the Artificial Analysis leaderboard for the current ranking before you decide.
Should I switch from OpenAI or Anthropic models to Microsoft's MAI models?
Not yet, and not wholesale. The honest answer is to test MAI Code 2 in GitHub Copilot on your own everyday coding tasks — it is the only model in the family that is generally available to most users right now — and keep your frontier models for the hard problems. The price claims are attractive but partly promotional, the independent benchmarks do not exist yet, and several models are still in preview or limited access. Microsoft itself is not asking you to choose: GPT-6.2 and Meta's latest models are also coming to Foundry, so the sensible setup is MAI models for high-volume routine work and frontier models where quality matters most.