Microsoft MAI models are the AI models Microsoft builds in house at Microsoft AI (MAI), the division led by Mustafa Suleyman, rather than licensing from OpenAI. At Ignite on November 17, 2026 Microsoft released seven of them: 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. The short version: they are specialist models priced well below the flagship models from OpenAI and Anthropic, and most of them are something you reach through a Microsoft product (GitHub Copilot, Microsoft Foundry, MAI Labs, Dragon Copilot) rather than a chat app of their own.
Microsoft is still OpenAI's biggest partner, so the new family is its clearest sign yet that it wants its own frontier lab, competing on price per task as much as on top scores. Below: every model, its price, where to use it, and which claims to read with care.
Microsoft MAI models at a glance
| Model | What it does | Where you can use it (Nov 20, 2026) | Published price |
|---|---|---|---|
| MAI Code 2 | Everyday enterprise coding, debugging, multifile changes | GitHub Copilot Pro, Pro+, Max, Business, Enterprise | $0.50 input, $2.50 output per 1M tokens (promo to May 17, 2027) |
| MAI Code 2 Flash | Fast, cheap routine coding | GitHub Copilot (including Free and Student); Foundry limited preview | $0.05 input, $0.25 output per 1M tokens (promo) |
| MAI Cyber 2 Flash | Vulnerability discovery and security investigation | Microsoft security products such as Codename MDASH, from mid December | $0.10 input, $0.50 output per 1M tokens (list price Microsoft used in testing) |
| MAI Image 2.7 and 2.7 Flash | Image generation and precise editing | Foundry public preview, MAI Labs, fal, Genspark, OpenRouter, Vercel | About $35 and $10 per 1,000 images |
| MAI Realtime 1 | Live speech to speech voice agents | MAI Labs; Foundry limited preview | From $0.01 per audio minute, reasoning billed separately |
| MAI Transcribe 2 | Multilingual transcription | Inside Dragon Copilot Transcribe | Not published separately |
| Decision 1.1 | Routes requests so pricier models run only when needed | Microsoft customers (about 4,000, per Microsoft) | Not published |
Microsoft also said its second generation AI chip, Maia 200, is now in production running its largest models, and it announced MAI Licensed Weights, a way for large companies to run and customize MAI models inside their own Azure environment.
What is Microsoft AI (MAI)?
Microsoft AI is the group Suleyman has run since 2024. Two changes turned it from a Copilot product team into a model lab. The October 2025 Microsoft and OpenAI agreement kept OpenAI as Microsoft's frontier model partner but let Microsoft pursue AGI on its own, and in November 2025 Suleyman set up the MAI Superintelligence Team around what he calls humanist superintelligence. Our guide to Microsoft AI Frontiers covers that reorganization and the research lab it absorbed.
The first full MAI model family arrived on June 2, 2026: MAI Thinking 1 (a reasoning model), MAI Code 1 Flash, MAI Image 2.5, MAI Transcribe 1.5 and MAI Voice 2. Microsoft said then that it trains these models without distilling other labs' models. The November release is the second wave. Suleyman's message in his Ignite keynote and to Reuters was the same: Microsoft improves quality and price together, with what he calls a "hill climbing machine."
The seven Ignite 2026 models, one by one
MAI Code 2 and MAI Code 2 Flash
These are the headline releases. Microsoft's Code 2 announcement says MAI Code 2 is built for the coding work customers do every day: debugging, multifile implementation, reasoning across a repository and building user interfaces. Averaged over four benchmarks (DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1), Microsoft says it reaches 98% of GPT-6.1 Sol's score at about 40% of the cost, and 96% of Claude Opus 5.5's score at 5% of the cost. In blind tests of frontend work, developers preferred its output over GPT-6 Sol 64% of the time.
MAI Code 2 Flash is the cheap, fast sibling for code explanations, targeted edits and routine fixes. Microsoft compares it with GPT-6 Luna and Claude Haiku 5.5 and says it performs comparably at 65% and 20% of their cost. If you are choosing between those two small models, see Claude Haiku 5.5 vs GPT-6 Luna.
The GitHub changelog confirms both models went live in GitHub Copilot on November 17. MAI Code 2 is on the Pro, Pro+, Max, Business and Enterprise plans; Code 2 Flash is also on Copilot Free and Student. GitHub's models and pricing page lists the usage based rates:
| Model | Input per 1M tokens | Cached input | Output per 1M tokens | After May 17, 2027 |
|---|---|---|---|---|
| MAI Code 2 | $0.50 | $0.025 | $2.50 | $1.00 / $0.05 / $5.00 |
| MAI Code 2 Flash | $0.05 | $0.005 | $0.25 | $0.10 / $0.01 / $0.50 |
Two caveats come from Microsoft's own footnotes. The cost comparisons use the 50% promotional price, so the gap shrinks once the promotion ends. And Microsoft says the models are "explicitly not yet tuned" for fundamental science research or very long agentic work, which is where flagship models like Opus 5.5 still earn their price. A model tuned from Code 2 Flash, MAI Work Flash, now serves Microsoft 365 apps such as Excel, Word and Teams.
MAI Cyber 2 Flash
MAI Cyber 2 Flash is a specialist security model. On the CyberGym-E2E vulnerability discovery benchmark Microsoft reports 92.3% task success against 91.5% for GPT-6 Astra, at less than 2% of the cost per task. It has a one million token context window, four times the size of MAI Cyber 1.1 Flash.
The benchmark note is unusually candid. Microsoft found that 92 of the benchmark's 905 tasks drew on the same open source projects and vulnerability reports as some of its training data, so it dropped those 92 tasks for every model and reported results on the other 813. You cannot buy this model as a general API: it is coming only to Microsoft security products, including Codename MDASH, where it replaces the existing MAI model for eligible customers from mid December.
MAI Image 2.7 and MAI Image 2.7 Flash
MAI Image 2.7 took first place on the Artificial Analysis image editing leaderboard as of November 16, at 1,215 Elo, 4 points ahead of OpenAI's GPT Image 2.6 Sunburst (max). Microsoft prices it at about $35 per 1,000 images and the Flash version at about $10 per 1,000. Both are in public preview on Microsoft Foundry and through fal, Genspark, OpenRouter and Vercel, and both can make transparent backgrounds. A new Canvas workspace in MAI Labs, powered by Flash, lets you select part of an image to replace, move or restyle. MAI Image also powers Bing Image Creator by default.
MAI Realtime 1
MAI Realtime 1 is Microsoft's first live speech to speech model, aimed at voice agents for customer service and similar work. Base usage starts at $0.01 per audio minute, with reasoning billed separately at $1 per million input tokens and $5 per million output tokens. It works with the OpenAI Realtime API format, which makes it easier to test against an existing voice app. You can try it in MAI Labs or sign up for the Foundry limited preview, and preview customers are not billed until public preview begins. Microsoft says it beat GPT Realtime 2.1 on its evaluations, including Big Bench Audio and Tau-Voice.
MAI Transcribe 2 and Decision 1.1
MAI Transcribe 2 powers Dragon Copilot Transcribe, Microsoft's clinical documentation tool, which Microsoft says handles 10 million patient encounters a month with multilingual error rates halved. Decision 1.1 is a routing model: it decides which model should answer a request so the expensive ones run only when needed. Microsoft says it is 50 times faster and 79 times cheaper than GPT-6.1 Sol, but publishes no price.
MAI Licensed Weights: Microsoft's answer to open weights
The most strategic announcement may not be a model at all. MAI Licensed Weights lets an organization deploy MAI models in its own Azure environment, adapt them with its own data, and decide who can use them and when they are upgraded. Microsoft says customer data and adaptations are used to train its models only with the customer's express agreement.
Microsoft cites early customers such as AT&T, whose customized model for 5G network analytics beat GPT-6.1 Sol on standard tasks at 1/50th of the cost, and Alaska Airlines, whose maintenance report classification accuracy rose from 81% to 91%. Licensed Weights is "coming soon" through Microsoft's Frontier Company engagements, so for now it is a conversation with a Microsoft account team, not a download.
How MAI models compare with OpenAI, Anthropic and Google
Microsoft is not claiming MAI has the smartest model. Its pitch is close to flagship quality on common work at a fraction of the price, which puts MAI against other labs' middle tiers as much as their flagships.
| Question | Microsoft MAI | OpenAI | Anthropic |
|---|---|---|---|
| Strategy | Specialist models per job, priced low | General flagship plus smaller tiers | General flagship plus smaller tiers |
| Main coding model | MAI Code 2 | GPT-6.1 Sol family | Claude Opus 5.5 |
| Where developers reach it | GitHub Copilot, Foundry, partners | OpenAI API, ChatGPT | Anthropic API, Claude |
| Consumer chat app | None of its own for these models | ChatGPT | Claude |
| Flagship weights in your own cloud | Licensed Weights (coming soon) | Not offered | Not offered |
For the OpenAI side, our GPT-6 Sol vs Astra vs Luna guide explains the tiers Microsoft benchmarks against. For the wider rivalry, see OpenAI vs Anthropic.
What about Microsoft Copilot, the assistant? Microsoft has not published which model answers each Copilot chat, and Copilot still offers OpenAI models. The Ignite news touches Copilot users indirectly: MAI Work Flash in Microsoft 365 apps, MAI Image in Bing Image Creator, and MAI Code models in GitHub Copilot. Our ChatGPT vs MS Copilot comparison covers the assistant itself.
Pros and cons of the MAI lineup
Pros
- Low published prices, especially Code 2 Flash, Image 2.7 Flash and Realtime 1.
- Easy to try if you already pay for GitHub Copilot or use Microsoft Foundry.
- Microsoft disclosed benchmark contamination on CyberGym-E2E and adjusted for it.
- Licensed Weights gives large companies a path to control and customize a frontier lab's model.
- Realtime 1 speaks the OpenAI Realtime API format, so switching costs are lower.
Cons
- Almost every performance claim is Microsoft's own testing, often on its internal MSBench platform.
- The coding cost comparisons use a promotional price that doubles after May 17, 2027.
- MAI Cyber 2 Flash, Transcribe 2 and Decision 1.1 are locked inside Microsoft products.
- Code 2 is not tuned for long agentic work or research grade problems.
- No consumer chat app of its own.
Should you use MAI models?
If you use GitHub Copilot, try MAI Code 2 for a week of ordinary tasks and compare results and credit spend with your usual model. If you build image features, MAI Image 2.7 Flash at about a penny per image is worth a test against what you pay now. If you run voice agents on the OpenAI Realtime API, MAI Realtime 1 is the easiest switch to trial.
If your work is hard research, long autonomous runs or anything where the top score matters more than cost, Microsoft itself says MAI Code 2 is not built for that yet. Stay with a flagship model and recheck MAI's next release. Treat the headline percentages as claims until independent leaderboards confirm them, a point an analyst made to InfoWorld about the promotional pricing.
FAQ
What are Microsoft MAI models?
They are AI models trained in house by Microsoft AI, Mustafa Suleyman's division, separate from the OpenAI models Microsoft also licenses. The family covers coding, security, images, voice, transcription and model routing, and it was expanded with seven releases at Ignite on November 17, 2026.
How much do MAI models cost?
MAI Code 2 costs $0.50 per million input tokens and $2.50 per million output tokens in GitHub Copilot until May 17, 2027, then $1 and $5. Code 2 Flash is $0.05 and $0.25, then $0.10 and $0.50. MAI Image 2.7 is about $35 per 1,000 images (Flash about $10), and MAI Realtime 1 starts at $0.01 per audio minute.
Are MAI models available in GitHub Copilot?
Yes. MAI Code 2 is on Copilot Pro, Pro+, Max, Business and Enterprise, and MAI Code 2 Flash is also on Copilot Free and Student. Business and Enterprise admins control access through the Copilot model policy.
Is MAI Code 2 better than Claude Opus 5.5?
Not by Microsoft's own numbers. It says MAI Code 2 reaches 96% of Opus 5.5's score on four coding benchmarks at 5% of the cost, measured at the promotional price. It is the cheaper choice for routine coding, while Opus 5.5 remains stronger on the hardest and longest tasks.
Can I download MAI model weights?
Not today. MAI Licensed Weights, announced at Ignite, will let organizations run and customize MAI models in their own Azure environment, but Microsoft says it is coming soon through its Frontier Company engagements, and access goes through a Microsoft account team.



