Microsoft's MAI unit has released seven new AI models that the company says deliver top-tier performance at a fraction of the cost of competing systems. The pitch is aimed at customers who want to run AI agents nonstop, a goal that gets expensive quickly as token usage rises.
The lineup spans cybersecurity, coding, image generation, speech and transcription, and model routing. Microsoft says the releases are meant to let developers and enterprises "run agents 24-7 doing real engineering work" without paying premium prices for every request.
The cybersecurity model, MAI Cyber 2 Flash, is the headline claim in the security space. Microsoft says it beats the quality of GPT-6 Astra at less than 2% of the cost per task on the CyberGym-E2E benchmark. The company argues that defenders now face attackers who have cheap and abundant offensive tools, and that monitoring and patching systems around the clock requires models that are affordable enough to run continuously.
Microsoft said the model builds on earlier MAI Cyber models. MAI Cyber accounts for almost 40% of scans in MDASH, the company said, and 90% of scans that started with MAI continued to use it. The latest model is priced 86% lower on output tokens than MAI Cyber 1.1 Flash, which Microsoft says gives existing MDASH customers a stronger model at a lower price.
Coding models aimed at everyday enterprise work
MAI Code 2 is designed for everyday enterprise programming tasks. Microsoft averaged results across four widely used benchmarks that reflect the coding work seen in GitHub Copilot: DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1. On that average, the model delivers 98% of GPT-6.1 Sol's performance at 40% of the price, and 96% of Claude Opus 5.5's performance at 5% of the price.
The company was clear about the model's limits. MAI Code 2 is not yet tuned for harder problems such as fundamental science research or very long-horizon agentic work. In a blind, side-by-side test of frontend work, developers preferred MAI Code 2 over GPT-6 Sol 64% of the time, Microsoft said. The model is available in GitHub Copilot.
MAI Code 2 Flash targets high-volume tasks such as code explanations, targeted edits, routine fixes and rapid iterations. Microsoft says it performs comparably to GPT-6 Luna and Claude Haiku 5.5 at 65% and 20% of their prices, respectively. It is available in GitHub Copilot and in a limited-access preview in Microsoft Foundry.
Microsoft also made claims about its image and speech tools. MAI Image 2.7 ranks first on Artificial Analysis for image editing, outperforming GPT Image 2.6 Sunburst (max) at one-fifth of the price. For text-to-image generation, the company says the model ranks fourth, which it describes as the second-best lab. A lower-cost version, MAI Image 2.7 Flash, is priced at about $10 per 1,000 images for high-volume work. Both are available in Microsoft Foundry.
MAI Realtime 1 is Microsoft's first live speech-to-speech model, built for natural, low-latency conversation. The company said it outperformed GPT-Realtime 2.1 on external and Dynamics 365 benchmarks. Pricing starts at $0.01 per audio minute. Customers can try it in MAI Labs.
Dragon Copilot Transcribe, which runs on MAI Transcribe 2, is now live in Dragon Copilot. Microsoft says it transcribes 100% of doctor-patient conversations in the product, covers 10 million encounters a month, and halves error rates in multilingual workflows. The company says the tool is used by a third of US physicians.
Decision 1.1 handles model routing inside an application, so larger and more expensive models are called only when needed. Microsoft says the model is 50 times faster and 79 times cheaper than GPT-6.1 Sol, and that it is used by 4,000 customers.
Custom models under customer control
Microsoft is also offering customers a way to tailor these models. The company said it has heard from business leaders who do not want to build on someone else's model. Through MAI Licensed Weights, customers can take a frontier MAI model, customize it with their own proprietary data if they choose, and run it in their own Azure environment.
Microsoft said customers control who can access the model, and that their data and adaptations remain theirs. The company will work with customers through Frontier Company to build the custom model and embed it in their products and agents. Microsoft said AT&T developed an MAI custom model for 5G network operation analytics that outperformed GPT-6.1 Sol at 50 times lower cost.
The announcement is a clear bid for enterprise customers who are weighing cost against capability. Microsoft's pitch rests on its own benchmark results, many of which were run by the company itself, so independent testing will be the real measure of how the models compare once customers put them to work.
This article was written with the assistance of AI.
News Factory APP - agentic news to boost your SEO & AEO.