Microsoft used its Ignite 2026 conference on Tuesday, November 17, to launch seven new first-party models from its Microsoft AI (MAI) group, led by the MAI Code 2 coding family, and to introduce MAI Licensed Weights, an offer that lets selected enterprises take an MAI model, tune it on their own data and run it inside their own Azure environment. It is Microsoft’s clearest bid yet to compete on price with OpenAI and Anthropic.
The seven models and where they run
Every performance and price comparison below comes from Microsoft itself:
- MAI Code 2, for everyday coding. Averaged over four benchmarks, Microsoft says it reaches about 98% of GPT-6.1 Sol’s score at roughly 40% of the price, and 96% of Claude Opus 5.5’s at about 5%, using a 50% promotional rate. It is live in GitHub Copilot.
- MAI Code 2 Flash, a cheaper model for routine edits and fixes, in GitHub Copilot and a limited-access preview in Microsoft Foundry.
- MAI Cyber 2 Flash, a security model with a one-million-token context window, list-priced at $0.10 per million input and $0.50 per million output tokens, used only inside Microsoft security products such as MDASH.
- MAI Image 2.7 and MAI Image 2.7 Flash, in public preview on Foundry at about $35 and $10 per 1,000 images respectively.
- MAI Realtime 1, Microsoft’s first speech-to-speech voice model, from $0.01 per audio minute, in MAI Labs and a limited Foundry preview.
- Dragon Copilot Transcribe, a medical transcription model built on MAI-Transcribe-2.
Microsoft says MAI Code 2 is not yet tuned for science research or very long agent runs. No benchmark figure has been checked independently.
What MAI Licensed Weights offers
Under the weights offer, customers can take a frontier MAI model, optionally adapt it with proprietary data, and deploy it in their own Azure environment, deciding who can access it and when it is upgraded. Microsoft says customer data and adaptations stay with the customer.
Access is not self-service: Microsoft builds each custom model with the customer through its Frontier Company engagements. It has not published prices, eligible models or licence terms, nor said whether weights can leave Azure.
Early users report their own results. AT&T tuned MAI Code 2 Flash and Code 2 for 5G network analytics at a fiftieth of GPT-6.1 Sol’s cost on standard tasks, and Alaska Airlines lifted maintenance-report classification accuracy from 81% to 91%. In October, SiliconANGLE reported Microsoft’s plan to license its MAI-Base-2 base model weights to trusted partners.
Licensed weights vs open weights and API-only models
Licensed weights are not open weights. Our guide to open-weight vs closed AI models covers the basic trade-off; here is where Microsoft’s new tier fits, based on what it has disclosed:
| Open weights | MAI Licensed Weights | API-only | |
|---|---|---|---|
| Who can get the weights | Anyone who accepts the licence | Selected enterprises, by agreement | Nobody outside the vendor |
| Where the model runs | Any hardware, cloud or on premises | The customer’s own Azure environment | The vendor’s servers |
| Customization | Full fine-tuning by anyone | Tuning on own data, built with Microsoft | Limited to vendor tuning tools |
| Upgrade timing | User decides | Customer decides | Vendor decides |
| Terms and price | Public licence, often free | Not published | Public per-token price list |
Licensed weights give a large company much of the control of an open-weight model, such as fixed versions and its own access rules, while Microsoft keeps the model in its cloud and under contract.
Copilot and agents, briefly
Satya Nadella framed Copilot as a “new OS for work”. Per Microsoft’s Book of News, Office in Copilot and Microsoft IQ are generally available, while the Autopilot agent and Copilot Workspaces go to Frontier program customers. Nadella said Agent 365 manages over 50 million agents. Ignite came days after a Microsoft 365 outage that hit Copilot.
What to watch next
- Licence terms, eligible models and pricing for MAI Licensed Weights.
- When MAI Code 2’s 50% promotional price ends.
- MAI Cyber 2 Flash replacing the current model in eligible MDASH profiles from mid-December.
Key takeaways
- Microsoft launched seven MAI models at Ignite 2026, led by MAI Code 2, now in GitHub Copilot.
- MAI Licensed Weights lets selected firms tune MAI models and run them in their own Azure environment.
- Licensed weights are not open weights: access is by agreement, terms are unpublished, and deployment is on Azure.
- All benchmark and cost claims so far come from Microsoft or its customers.
Frequently asked questions
Are Microsoft’s MAI models open source?
No. MAI models are offered through GitHub Copilot, Microsoft Foundry and Microsoft products, and weights go only to selected enterprises under licence, not as a public download.
How much does MAI Code 2 cost?
Microsoft has not published per-token prices for MAI Code 2 on its launch page. It says the model costs about 40% of GPT-6.1 Sol and about 5% of Claude Opus 5.5 at a 50% promotional rate.
How can a company get MAI Licensed Weights?
Microsoft says access runs through its Frontier Company engagements, and interested organizations should contact their Microsoft account team. No public sign-up, price or general availability date has been announced.



