Microsoft's Seven New MAI Models: Cheaper AI Takes on Rivals

Published (Updated )

Microsoft Just Bet Big on Cheaper, Specialist AI
Microsoft has unveiled seven new first-party AI models, and the message is hard to miss: top-tier AI does not have to come with a top-tier price tag.
- The models were announced at Microsoft Ignite 2026 on 17 November 2026 in San Francisco.
- They were introduced by Microsoft AI, the division led by Mustafa Suleyman.
- The pitch is near-frontier quality at a fraction of what rival models cost.
- Reuters confirmed the launch and reported that Microsoft is positioning these domain-specific models directly against OpenAI and Anthropic.
- The range covers coding, cybersecurity, image generation, real-time voice, transcription and smart model routing.
This is not a single general-purpose chatbot. It is a toolkit of specialist models, each built for one job and priced to undercut the big generalists. Here is the quick answer to what matters most:
- MAI Code 2 is the flagship coding model, now generally available in GitHub Copilot.
- MAI Code 2 Flash is a lighter, cheaper coding model for everyday programming work.
- MAI Cyber 2 Flash is a defensive security model with a huge 1-million-token context window.
- MAI Image 2.7 handles image generation and editing, with a cheaper Flash version alongside it.
- MAI Realtime 1 powers real-time voice conversations.
- MAI Transcribe 2 handles transcription and already supports a major healthcare product.
- Decision 1.1 is a routing model that picks the right model for each request.
The rest of this article breaks down each model, what Microsoft claims it can do, what it costs, where you can use it, and β just as importantly β where the claims deserve a closer look.
The Seven Models at a Glance
Here is the full line-up in one place, so you can see how the family fits together:
- MAI Code 2 β agentic coding, the flagship. Generally available in GitHub Copilot.
- MAI Code 2 Flash β everyday coding at a lower price. Live in Copilot, limited preview in Foundry.
- MAI Cyber 2 Flash β defensive cybersecurity. Limited availability, rolling out to eligible customers from mid-December.
- MAI Image 2.7 β image generation and editing. Public preview in Microsoft Foundry.
- MAI Realtime 1 β real-time voice. Available to try in MAI Labs.
- MAI Transcribe 2 β transcription, powering Dragon Copilot.
- Decision 1.1 β routing model that chooses which model handles a request. Generally available.
There is also an MAI Image 2.7 Flash variant for cheaper, faster image work. And sitting alongside the models is MAI Licensed Weights, a programme that lets selected enterprises customise the models with their own proprietary data and run them inside their own Azure environments.
The strategy is clear. Instead of asking one giant model to do everything, Microsoft is building smaller, focused models that do one thing well β and charging much less for them.
MAI Code 2: The Flagship Coding Model
MAI Code 2 is the headline act. It is an agentic coding model, which means it is designed to work on coding tasks over longer stretches, planning steps, running tools and producing working code rather than just completing a line or two.
- It is generally available now in GitHub Copilot.
- Microsoft claims it reaches 98% of GPT-6.1 Sol's score at around 40% of its price across four coding benchmarks.
- Microsoft also claims it delivers 96% of Claude Opus 5.5's performance at 5% of the price.
- Those are striking numbers β if they hold up in real-world use.
This is where the caveats matter, and Microsoft's own documentation gives us several:
- The price comparison uses a 50% promotional rate, so the real-world price gap may be smaller once that promotion ends.
- The benchmarks are Microsoft's chosen tests, not an independent or industry-wide standard.
- The model card says generated code can be incorrect, insecure or outdated, so human review remains essential.
- Quality can degrade in long agent sessions, which matters for the long, multi-step tasks agentic coding is meant for.
- The model is not tuned for fundamental science research or very long-horizon agentic work.
- Performance outside English and less common programming languages has not been extensively tested.
None of that makes MAI Code 2 a bad model. It makes it a model whose headline numbers should be read as Microsoft's best-case framing, not as a guarantee. If a team can run a capable coding agent for a fraction of the cost of the most expensive rivals, that opens the door to smaller companies, independent developers and larger teams running more agent hours.

MAI Code 2 Flash: Everyday Coding, Lower Cost
Not every coding task needs the flagship. MAI Code 2 Flash is the lighter sibling, aimed at everyday programming work where speed and cost matter more than maximum capability.
- It is live in GitHub Copilot now.
- It is also available in limited preview in Microsoft Foundry.
- Microsoft claims it matches GPT-6 Luna at 65% of the price.
The idea is simple: routine code completion, small fixes and everyday edits do not need the most powerful model in the line-up. A cheaper, faster model that handles those jobs well keeps costs down, especially for teams running AI coding tools across dozens or hundreds of developers. For most developers, this may end up being the model they use most often β the flagship gets the headlines, but Flash could quietly do the bulk of the daily work.
MAI Cyber 2 Flash: A Security Specialist
Cybersecurity is the most specialised model in the family, and arguably the most interesting. MAI Cyber 2 Flash is a defensive security model, built to help security teams find and fix vulnerabilities rather than to attack systems.
- It has a 1-million-token context window, which means it can take in very large amounts of code and security data at once.
- Pricing is $0.10 per million tokens for input and $0.50 per million tokens for output.
- Microsoft claims 92.3% task success on the CyberGym-E2E benchmark, versus 91.5% for GPT-6 Astra, at under 2% of the cost per task.
That cost claim is dramatic, but it comes with an important caveat: some of the benchmark's 905 tasks came from the same open-source projects and vulnerability reports as some of the model's training data. That overlap can inflate a benchmark score, so the result should be treated with caution until independent testing confirms it.
Availability is also limited. MAI Cyber 2 Flash is not on general sale. It is available to a limited group of customers, and it is rolling out to eligible Codename MDASH customers from mid-December.
MAI Image 2.7: Image Generation and Editing
MAI Image 2.7 is Microsoft's image model, covering both generation and editing. It is in public preview in Microsoft Foundry.
- Microsoft says it ranks number one for image editing on Artificial Analysis, a well-known independent benchmark.
- Pricing is around $35 per 1,000 images.
- Microsoft claims Image 2.7 beats GPT Image 2.6 Sunburst (max) at one-fifth of the price.
- The Flash variant, MAI Image 2.7 Flash, costs around $10 per 1,000 images.
Image editing is a useful thing to lead on. Many AI image tools are good at generating pictures from scratch but weaker at editing an existing image β changing an object, fixing a detail or restyling a scene. If Microsoft's top ranking for editing holds up, that is a genuine selling point for designers, marketers and product teams.
The pricing also matters. At around $35 per 1,000 images for the full model and around $10 per 1,000 for Flash, image generation becomes cheap enough to use at scale β for product mock-ups, marketing variations and design exploration.

MAI Realtime 1: Real-Time Voice
MAI Realtime 1 is the voice model in the family, built for live, low-latency conversations with an AI.
- Pricing starts from $0.01 per audio minute.
- It is available to try in MAI Labs, Microsoft's experimental playground.
Real-time voice is becoming a key battleground. Voice agents for customer support, hands-free assistants and interactive apps all need a model that can listen, understand and respond without awkward delays. At $0.01 per audio minute, the starting price is low enough for developers to experiment, though real-world costs will depend on how much voice traffic an app actually handles. Being in MAI Labs rather than general availability suggests Microsoft is still gathering feedback and refining the model.
MAI Transcribe 2: The Quiet Workhorse
MAI Transcribe 2 is the transcription model, and it may already be the most widely used model in the family β even though it is the least flashy.
- It underpins transcription in Dragon Copilot, Microsoft's healthcare AI assistant.
- Microsoft reports it supports 10 million patient encounters per month.
That number tells you a lot. Transcription in healthcare has to be accurate, fast and reliable, because clinicians depend on it to document patient visits. Handling 10 million encounters a month is a serious real-world workload, and it gives Microsoft a strong proof point that the model works at scale in a demanding setting. It also shows Microsoft's strategy in action: build a specialist model, prove it in a high-stakes product, and then offer it more widely.
Decision 1.1: The Traffic Controller
Decision 1.1 is the least glamorous model in the line-up, but it may be one of the most important. It is a routing model β its job is to look at a request and decide which model should handle it.
- It is generally available now.
- It sits in front of the other models and routes each task to the best fit.
Why does this matter? Because a family of specialist models only works if the right model gets each job. A routing model that sends simple tasks to a cheap model and hard tasks to a powerful one is how you keep quality high and costs low. Without it, users would have to guess which model to use every time. Decision 1.1 is the piece that makes the whole family usable as a system.
MAI Licensed Weights: Custom Models for Big Enterprises
Alongside the seven models, Microsoft announced MAI Licensed Weights. This programme lets selected enterprises customise the models using their own proprietary data and deploy them inside their own Azure environments β a direct play for large organisations with strict data and compliance needs that do not want to send sensitive data to a shared public model.
What This Means for OpenAI and Anthropic
Reuters reported that Microsoft is positioning these domain-specific models against OpenAI and Anthropic, and the pricing claims make that clear. Microsoft is not trying to beat its rivals on every possible task. It is trying to win on value: close to the same quality, for much less money, in the specific jobs that matter most.
- For developers, cheaper capable coding models could cut the cost of AI-assisted development significantly.
- For businesses, specialist models for security, transcription and image work could replace more expensive general-purpose models for those tasks.
- For OpenAI and Anthropic, the pressure is on pricing. If Microsoft's models are genuinely close in quality at a fraction of the cost, the premium for the biggest models becomes harder to justify.
- For the wider AI market, this could accelerate a shift from one giant model to many specialist models.
The MAI models are also a statement of independence: Microsoft is showing it can build its own competitive models, not just rely on someone else's technology.
Pricing Snapshot
Here are the prices Microsoft has disclosed, all in US dollars:
- MAI Cyber 2 Flash: $0.10 per million input tokens, $0.50 per million output tokens.
- MAI Image 2.7: around $35 per 1,000 images.
- MAI Image 2.7 Flash: around $10 per 1,000 images.
- MAI Realtime 1: from $0.01 per audio minute.
- MAI Code 2 and MAI Code 2 Flash: pricing is framed as a percentage of rival models' prices, with the flagship claimed at around 40% of GPT-6.1 Sol's price and Flash at 65% of GPT-6 Luna's price.
Remember that the MAI Code 2 price comparison uses a 50% promotional rate, so the gap may narrow once it ends.
The Caveats to Keep in Mind
Microsoft's launch is genuinely significant, but the claims deserve the usual scrutiny. Here is what to keep in mind:
- All benchmark and price comparisons are Microsoft's own claims, based on Microsoft's chosen tests.
- The MAI Code 2 price comparison relies on a 50% promotional rate, so check the current price before budgeting.
- The MAI Cyber 2 Flash benchmark includes tasks drawn from the same sources as some of its training data.
- The MAI Code 2 model card warns that generated code can be incorrect, insecure or outdated, and that quality can degrade in long sessions.
- Performance outside English and less common languages has not been extensively tested.
- MAI Cyber 2 Flash is not generally available, and MAI Realtime 1 is still in MAI Labs.
None of these caveats cancel out the launch, and Microsoft deserves credit for publishing a model card that spells out the limitations. But they are a reminder that the real test is independent evaluation and everyday use, not launch-day numbers.
Where You Can Try the Models
Availability varies by model, so here is a quick guide:
- Generally available now: MAI Code 2 in GitHub Copilot, and Decision 1.1.
- Live in Copilot, limited preview in Foundry: MAI Code 2 Flash.
- Public preview in Foundry: MAI Image 2.7.
- Try in MAI Labs: MAI Realtime 1.
- Already in a live product: MAI Transcribe 2, inside Dragon Copilot.
- Limited group, rolling out from mid-December: MAI Cyber 2 Flash, for eligible Codename MDASH customers.
The Bottom Line
Microsoft's seven MAI models are a serious bet on a different shape for the AI market: specialist models, priced aggressively, aimed at the jobs where most of the real work happens.
- The coding models are the most immediately useful, and they are already in GitHub Copilot.
- The image and voice models bring capable AI to creative and conversational work at low prices.
- The security and transcription models show Microsoft targeting high-value, high-trust use cases.
- Decision 1.1 and MAI Licensed Weights show this is a platform play, not a one-off launch.
The claims are bold and the caveats are real. The next few months will tell us whether the models live up to their billing once developers, security teams and designers put them through their paces. But one thing is already clear: the era of paying top dollar for a general-purpose model to do a specialist's job is coming under serious pressure β and Microsoft just made that pressure a lot harder to ignore.
FAQs
What are Microsoft's seven new MAI models?
The seven models unveiled at Ignite 2026 are MAI Code 2, MAI Code 2 Flash, MAI Cyber 2 Flash, MAI Image 2.7, MAI Realtime 1, MAI Transcribe 2 and Decision 1.1. They cover coding, cybersecurity, image generation and editing, real-time voice, transcription and model routing.
When were the MAI models announced?
Microsoft unveiled the seven MAI models at Ignite 2026 on 17 November 2026 in San Francisco. They were introduced by Microsoft AI, the division led by Mustafa Suleyman.
Is MAI Code 2 available in GitHub Copilot?
Yes. MAI Code 2 is generally available in GitHub Copilot, and the lighter MAI Code 2 Flash is also live in Copilot, with a limited preview in Microsoft Foundry.
How much do the MAI models cost?
Microsoft prices them aggressively: MAI Cyber 2 Flash costs $0.10 per million input tokens and $0.50 per million output tokens, MAI Image 2.7 costs around $35 per 1,000 images, the Flash image variant costs around $10 per 1,000 images, and MAI Realtime 1 starts from $0.01 per audio minute. Microsoft claims MAI Code 2 costs around 40% of GPT-6.1 Sol's price, but that comparison uses a 50% promotional rate.
Is MAI Cyber 2 Flash available to everyone?
No. MAI Cyber 2 Flash is not on general sale. It is available to a limited group of customers and is rolling out to eligible Codename MDASH customers from mid-December.
How good are the MAI models compared with rivals?
Microsoft claims near-frontier quality at a fraction of rivals' prices β for example, 98% of GPT-6.1 Sol's coding score at around 40% of the price for MAI Code 2. These are Microsoft's own claims based on its chosen benchmarks, and the MAI Code 2 price comparison uses a 50% promotional rate, so independent testing will be the real test.
Was this article helpful?

Written By
TechorCoffeeContributing Writer
TechorCoffee is a technology reporter and software creator writing reviews and tutorials for TechOrCoffee.
Comments (0)
Leave a Comment
Read Next

Tue Oct 13 2026
Gemini 4 Argon: Everything You Need to Know
Gemini 4 Argon is Googleβs new flagship AI model. See its benchmarks vs GPT-6 Astra and Claude, pricing, 1M-token output, safety features and how to get access.

Fri Oct 30 2026
Qwen 4 Is Not Out Yet: What Is Actually Confirmed
Qwen 4 in 2026: what Alibaba confirmed at Apsara, what the Flash-Next preview shows, what Kalmia rumours claim, and what is still unknown before launch.

Mon Oct 26 2026
Instagram Now Available on Apple TV: Check Out the Details
Social Media will be featured on your home screens! People can now scroll reels and watch videos in their living space with just one step.