Microsoft has spent 2026 building its own AI models, and at its Ignite conference on 17 November it announced a new wave of them. Microsoft AI’s “MAI” family now covers coding, cybersecurity, images, live speech, transcription and decision routing, and Microsoft says the newest models match rivals from OpenAI and Anthropic at a fraction of the cost. If you have seen “MAI” in the headlines and want a clear briefing, this is it.
Key takeaways
- MAI is Microsoft AI’s in-house model family. It grew through 2026 in four steps: April, June, August and the Ignite wave on 17 November.
- The performance and price comparisons are Microsoft’s own. Axios notes they rely on selected benchmarks and that some price advantages reflect promotional discounts.
- You can use some of it today, including MAI Code 2 in GitHub Copilot and MAI Image 2.7 in Foundry public preview. MAI Licensed Weights is only “coming soon”.
- This is not Microsoft leaving OpenAI. Microsoft still hosts OpenAI and Anthropic models, and MAI is its push for “long term self-sufficiency”.
Below we cover what MAI is, how it grew, what the new models claim, where you can try them and what is still unverified. We have not tested any of these models. Everything comes from Microsoft’s own posts or independent reporting, and we say which is which.
What MAI is
MAI is the model family built by Microsoft AI, the group led by CEO Mustafa Suleyman. Axios notes that Microsoft “has long relied on OpenAI and other AI companies for its most advanced models,” and that the new releases are meant to show it can build competitive technology of its own.
Microsoft says it trains its reasoning models from scratch and does not distill from other labs. That is Microsoft’s description of its own process, and we could not check it. Its stated goal, from the 2 June announcement: “This is all about long term self-sufficiency for Microsoft and our partners.”
That is not a break with partners. GeekWire reported in June that Microsoft’s Foundry platform also hosts the latest models from OpenAI and Anthropic, and that Microsoft has invested a cumulative $13 billion in OpenAI and announced up to $5 billion in Anthropic. Microsoft’s Ignite post does not say it is dropping any partner.
How the MAI family grew in 2026
| Date | What Microsoft announced | Details Microsoft gives |
|---|---|---|
| 2 April | MAI-Transcribe-1, MAI-Voice-1 and MAI-Image-2 in Foundry public preview | Transcribe-1 covers 25 languages; Voice-1 can make 60 seconds of audio in under a second on one GPU; Image-2 debuted in the top three image model families on Arena.ai |
| 2 June (Build) | “Seven new models”: Thinking 1, Code 1 Flash, Image 2.5 (plus a Flash variant), Transcribe-1.5 and Voice 2 | Code 1 Flash has 5 billion active parameters; Transcribe-1.5 supports 43 languages; Voice 2 supports 15+ languages |
| 12 August | MAI Thinking 1 in public preview on Foundry | 35B active, about 1T total parameters, sparse mixture-of-experts; 256k-token context |
| 17 November (Ignite) | Code 2, Code 2 Flash, Cyber 2 Flash, Image 2.7, Realtime 1, Transcribe 2, Decision 1.1 and Maia 200 chips | Covered below |
Microsoft says Thinking 1 was preferred over Anthropic’s Claude Sonnet 4.6 in a blind human comparison of 1,276 tasks run with an outside partner, Surge. That is a Microsoft-run evaluation. GeekWire described Thinking 1 as a private preview in June; Microsoft’s 12 August update says public preview.
What Microsoft announced at Ignite
Microsoft’s post says it is releasing “a family of seven new world class models,” then lists eight bullets, one of which is a chip. Axios counted “six new AI models.” We do not pick a single count, because the sources disagree.
The count depends on who is counting
7models, in Microsoft’s headline
8bullets in Microsoft’s list, including a chip
6models, according to Axios
MAI Code 2 and Code 2 Flash
Microsoft says that, averaged over four coding benchmarks it chose, MAI Code 2 gives “98% of the performance of GPT-6.1 Sol at 40% of the price,” and 96% of Claude Opus 5.5’s performance at 5% of the price. It says developers preferred Code 2 over GPT-6 Sol 64% of the time in a blind frontend comparison. Microsoft is upfront that the models are “explicitly not yet tuned for harder problems, like fundamental science research and very long horizon agentic work.” Code 2 Flash is the smaller option, which Microsoft says performs comparably to GPT-6 Luna and Claude Haiku 5.5 at 65% and 20% of their prices.
The Ignite post gives percentages, not per-token prices, so you cannot turn “40% of the price” into dollars from it. Axios also reports that OpenAI announced GPT-6.2 Astra and GPT-6.2 Sol the same day, so Microsoft’s comparisons name earlier versions. For a baseline on OpenAI’s model, see our GPT-6.1 Sol API quickstart.
MAI Cyber 2 Flash
Microsoft says it “beats GPT-6 Astra’s quality at less than 2% of the cost per task” on the CyberGym-E2E benchmark, and is priced 86% lower on output tokens than MAI Cyber 1.1 Flash. It is aimed at defense agents that run around the clock. The post does not say how outside customers get it, so we treat access as unconfirmed.
MAI Image 2.7
Microsoft says MAI Image 2.7 ranks #1 on Artificial Analysis image editing, ahead of GPT Image 2.6 Sunburst (max) at one-fifth of the cost. Its own footnote gives the margin: as of 16 November, 1,215 Elo across 5,478 votes, ahead by 4 Elo. That is a narrow lead on a leaderboard that keeps moving. For text-to-image generation, Microsoft says it ranks #4. It lists about $35 per 1,000 images for Image 2.7 and about $10 per 1,000 for the Flash version, and says MAI Image is now the default family behind Bing Image Creator.
MAI Realtime 1, Transcribe 2 and Decision 1.1
Microsoft calls Realtime 1 its first live speech-to-speech model and says it outperformed GPT-Realtime 2.1 in its own evaluations. Pricing starts at $0.01 per audio minute, but the post points to MAI Labs and a form to register interest.
Dragon Copilot Transcribe, built on MAI Transcribe 2, is live in Dragon Copilot. Microsoft says it covers 10 million patient encounters a month and halves multilingual error rates. It calls the model “the world’s fastest and most accurate multilingual transcription model,” a vendor claim with no outside benchmark in the post.
Decision 1.1 routes work so bigger, pricier models are used only when needed. Microsoft says 4,000 customers use it and that it is 50x faster and 79x cheaper than GPT-6.1 Sol. Maia 200, Microsoft’s second-generation accelerator, is in production with 1.7x more tokens per watt, though the post does not say against what.
What MAI Licensed Weights is
MAI Licensed Weights would let an organization deploy and customize Microsoft AI models in its own Azure environment. Microsoft says customer data and adaptations stay the customer’s, and that it uses them to train its models only with express agreement.
It is “coming soon through our Frontier Company engagements,” and interested customers are told to contact their Microsoft account team. It is not generally available, and it is not a download for a laptop: Axios reports deployment must happen inside Azure, and quotes Suleyman saying it would “feel like open source.” To run an open model on your own machine instead, see our guide to running an AI model locally with Ollama.
Microsoft’s customer examples come from task-specific tuning, not the general models. AT&T says a customized model beat GPT-6.1 Sol on standard 5G network analytics tasks at 1/50th of the cost. Alaska Airlines says its maintenance report classification accuracy rose from 81% to 91%, with qualified employees keeping responsibility for final decisions.
Where you can use MAI today
Access as Microsoft describes it. Prices are Microsoft’s listed figures and can change.
| Model | Where, according to Microsoft |
|---|---|
| MAI Code 2 | GitHub Copilot, available now |
| MAI Code 2 Flash | GitHub Copilot; limited-access preview in Foundry |
| MAI Image 2.7 and 2.7 Flash | Foundry public preview, plus fal, Genspark, OpenRouter, Vercel and MAI Labs |
| MAI Realtime 1 | MAI Labs, with a form to register interest |
| MAI Transcribe 2 | Inside Dragon Copilot |
| MAI Thinking 1 | Foundry public preview (listed at $2 and $8 per million input and output tokens at the Build launch) |
| MAI Voice 2, Transcribe-1.5 | Azure Speech (from $22 per million characters and $0.36 per hour at launch) |
| MAI Cyber 2 Flash, Decision 1.1 | Access not spelled out in the Ignite post |
| MAI Licensed Weights | Coming soon, through Frontier Company engagements |
Voice cloning is the high-risk part
MAI Voice 2 adds “identity preservation,” which Microsoft says recreates a specific person’s vocal identity so it can “speak as” them across languages. Microsoft pitches it for branded voices and accessibility, but it is also a capability that can be misused for impersonation. Microsoft says Voice 2 has “strong safeguards against misuse,” and in April said custom voices on its earlier Voice-1 model need an approval process aligned with its responsible AI policies. The posts we read do not detail the Voice 2 safeguards. If a voice message ever pushes you to act fast, our guide on how to spot a phishing message explains how to verify it another way.
Why Microsoft is building its own models
The reporting points to cost, control and ambition. Axios says MAI Code 2 in GitHub Copilot could shift “some coding workloads away from outside models,” though Suleyman declined to give a target. Suleyman also told Axios the company is making progress on its own frontier general-purpose model, with no release timeline. Separately, Microsoft is one of five companies named so far in an unconfirmed report that the US plans to designate AI “national champions”; see our explainer.
What it means for you
- If you use GitHub Copilot, try MAI Code 2 on your own repository and judge it there, not on a benchmark.
- If you build with images, check the current Foundry price for Image 2.7, since Axios says some price advantages are promotional.
- If you run a business, Licensed Weights is a conversation with a Microsoft account team, not a product you can buy yet. Ask about price, timing and what runs in your tenant.
- If you depend on one vendor, remember availability too. Our Microsoft 365 outage explainer covers a separate incident on 12 November when Outlook, Teams and Copilot went down.
What we don’t know yet
- The model count. Seven, eight or six depending on the source.
- Code 2 dollar prices. Only percentages are given.
- Access to Cyber 2 Flash and Decision 1.1.
- Licensed Weights pricing and timing.
- Independent testing. We found no outside evaluation of the new Code 2, Cyber 2 Flash or Realtime 1 claims. Leaderboard ranks come from third-party sites but are reported by Microsoft.
What to watch next
Watch for a published Code 2 price, wider Realtime 1 access, the first Licensed Weights customers and independent comparisons of the coding models. We will update this post when they arrive. For more model news, see our AI News section.
Frequently asked questions
What is Microsoft MAI?
MAI is the family of AI models built by Microsoft AI, Microsoft’s in-house model group. It covers coding, reasoning, images, voice, transcription, cybersecurity and decision routing. Its latest wave was announced at Ignite on 17 November 2026.
Is Microsoft replacing OpenAI with MAI?
Nothing Microsoft has published says so. GeekWire reported that Foundry also hosts the latest OpenAI and Anthropic models. Microsoft calls MAI a step toward “long term self-sufficiency,” and Axios says it could reshape the relationship, but no partner has been dropped.
Can I use MAI models today?
Some, yes. Microsoft says MAI Code 2 is available in GitHub Copilot, and MAI Image 2.7 is in public preview on Microsoft Foundry and through partners. MAI Licensed Weights is “coming soon.”
How good is MAI Code 2 compared with GPT-6.1 Sol?
Microsoft says it reaches 98% of GPT-6.1 Sol’s coding performance at 40% of the price, averaged over four benchmarks it chose. We have not tested it, and Microsoft says it is not tuned for very hard problems.
Quick quiz
What does Microsoft say MAI Licensed Weights lets you do?
- Download MAI models to a laptop
- Deploy and customize MAI models in your own Azure environment
- Replace GitHub Copilot
Microsoft says it lets organizations deploy and customize its models in their own Azure environments, and that it is coming soon.
How many new models did Axios say Microsoft announced at Ignite?
- Six
- Seven
- Eight
Axios described “six new AI models,” while Microsoft’s headline says seven and its list has eight bullets.
Sources
- Microsoft AI: Seven new world-class models at scalable price-performance (17 November 2026)
- Microsoft AI: MAI Image 2.7: Frontier quality at a fraction of the cost (17 November 2026)
- Microsoft AI: MAI Licensed Weights: Frontier performance you control (17 November 2026)
- Microsoft AI: Building a hill-climbing machine: Launching seven new MAI models (2 June 2026)
- Microsoft AI: Introducing MAI Thinking 1 (12 August 2026)
- Microsoft Community Hub: New MAI models in Microsoft Foundry (2 June 2026, updated 12 August)
- Microsoft Community Hub: Introducing MAI-Transcribe-1, MAI-Voice-1, and MAI-Image-2 in Microsoft Foundry (2 April 2026)
- Axios: Microsoft debuts six AI models as it pushes toward frontier (17 November 2026)
- GeekWire: Microsoft unveils seven homegrown AI models in new bid for ‘long term self-sufficiency’ (2 June 2026)





Write a response