Microsoft launches seven MAI models to lower enterprise AI costs
The new MAI lineup offers high-performance AI for coding, security, and images at a fraction of the price of current frontier models.

The short version
- MAI Code 2 reaches 98% of GPT-6.1 Sol quality at roughly 40% of the price.
- MAI Cyber 2 Flash offers a one-million-token context window for security tasks.
- MAI Image 2.7 ranks first on Artificial Analysis image editing leaderboards.
- Microsoft-Decision-1.1 routes requests to larger models only when necessary to save costs.
- MAI Licensed Weights allows enterprises to run customized models within their own Azure environments.
Why it matters
These models allow companies to run complex AI agents and workflows without the high costs typically associated with top-tier models. It signals a shift toward price-performance competition where specialized, cheaper models can replace general-purpose ones for specific tasks.
A new era of price-performance
Microsoft introduced a family of seven new MAI models at Ignite 2026. The primary goal of this lineup is to provide enterprises with frontier-level results without the high costs associated with standard frontier token bills. These models are designed for customers who need to run AI agents around the clock.
The release covers a wide variety of use cases including coding, security, image generation, voice, transcription, and model routing. By offering these specific tools, Microsoft aims to position itself as a lower-cost model supplier across its Copilot, Foundry, and Azure platforms. This strategy directly competes with OpenAI and Anthropic on the basis of price-performance.
The company is also introducing MAI Licensed Weights as part of this rollout. This feature allows enterprises to customize a frontier MAI model with their own private data. Once customized, these models can be run inside a company's own Azure environment for increased control and privacy.
Coding and cybersecurity tools
The coding side of the new lineup features MAI Code 2 and MAI Code 2 Flash. Microsoft reports that Code 2 reaches 98% of the quality of GPT-6.1 Sol while costing about 40% of the price. It is specifically tuned for everyday enterprise work and frontend output rather than science research or very long-horizon tasks.
Code 2 Flash is designed for routine edits and fixes. It matches GPT-6 Luna at 65% of the price and Claude Haiku 5.5 at 20% of the cost. Both of these coding models are already live in GitHub Copilot, with the Flash version also appearing in a limited preview on Microsoft Foundry.
For security, the company launched MAI Cyber 2 Flash. This model scores 92.3% on the CyberGym-E2E benchmark, which is higher than the 91.5% score of GPT-6 Astra. It achieves this performance at under 2% of the cost per task and includes a one-million-token context window.
Images and audio capabilities
MAI Image 2.7 has already gained attention for its performance in image editing. It currently ranks first on the Artificial Analysis image editing leaderboard, beating GPT Image 2.6 Sunburst at one fifth of the price. A Flash version is also available, which costs about $10 per 1,000 images compared to $35 for the full model.
These image models are available in public preview on Foundry and through various partners like fal, Genspark, OpenRouter, and Vercel. Additionally, the Flash model powers a new Canvas workspace within MAI Labs. The company's overall standing in image generation is now listed as the second-ranked lab across three leaderboards by Arena.
The lineup also includes audio and transcription tools. MAI Realtime 1 is Microsoft's first live speech-to-speech model, with interest registration now open through MAI Labs. For transcription, Dragon Copilot Transcribe uses MAI Transcribe 2 and is already processing 10 million patient encounters every month.
Smart routing and hardware
To manage costs efficiently, Microsoft released Microsoft-Decision-1.1. This is a routing model that decides whether to send a request to a larger, more expensive model or handle it with a smaller one. It received a score of 68 on the Decision Index from Artificial Analysis, an improvement over the previous version's score of 62.
This routing model is now generally available to help businesses optimize their AI spending. By only using high-powered models when necessary, companies can reduce their overall token bills while maintaining high performance. It serves as a foundational layer for managing multi-model workflows.
Supporting these models is the Maia 200 accelerator. Microsoft confirmed that this hardware is now in production to run its largest models. This infrastructure provides the necessary power to host the MAI family while maintaining the low-cost objectives of the software release.
What to watch
- How enterprises adopt MAI Licensed Weights for private data customization.
- The performance of MAI Realtime 1 as it moves from registration to public use.
- Whether the lower-cost MAI models begin to shift market share away from OpenAI and Anthropic.
Try it
Check the GitHub changelog to see which MAI models are currently supported in your specific Copilot plan.
Sources and links
First reported by testingcatalog.com. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.