On November 17, 2026, Microsoft unveiled seven new MAI models at Ignite 2026, covering coding, security, image, voice, and transcription. In its official blog, the Microsoft AI team published benchmark scores and prices against rival models one by one, and the intent is clear: the contest is not about parameter scale, but about driving down the cost of running agents around the clock.
Coding and security: flagship-level quality at a fraction of the price
MAI Code 2 is the flagship coding model of this release. Microsoft says that across four everyday enterprise coding benchmarks, it reaches 98% of GPT-6.1 Sol's performance at about 40% of the price; against Claude Opus 5.5, it scores 96% at just 5% of the price. The lighter Code 2 Flash matches GPT-6 Luna and Claude Haiku 5.5 at 65% and 20% of their prices, and both models are now in GitHub Copilot. On the security side, MAI Cyber 2 Flash is more aggressive: on the CyberGymE2E benchmark, it delivers comparable quality to GPT-6 Astra at less than 2% of the cost per task, with output tokens 86% cheaper than the previous generation — built so defense agents can patrol around the clock without worrying about the bill.
Image and voice: Microsoft is after two more territories
MAI Image 2.7 ranks first on Artificial Analysis's image-editing leaderboard; Microsoft says it beats the top configuration of GPT Image 2.6 Sunburst at one-fifth of the price. Its text-to-image ranking is fourth, and the lightweight version costs about $10 per 1,000 images. On the voice side, MAI Realtime 1, the first speech-to-speech model, starts at $0.01 per audio minute, and Microsoft says it outperformed GPT-Realtime 2.1 in its own evaluations. The transcription model MAI Transcribe 2 already powers Dragon Copilot, processing 10 million doctor-patient conversations a month and cutting multilingual error rates in half.
Beyond the models, Microsoft also introduced MAI Licensed Weights: companies can take a frontier MAI model, customize it with their own data, and run it in their own Azure environment, keeping access and data in their own hands. One caveat: all the comparison figures above come from Microsoft's own evaluation methodology, and independent third-party verification has yet to catch up. However large the price advantage, it still pays to test on real tasks at small scale before rolling out widely. The models Microsoft kept benchmarking against are OpenAI's GPT-6.2 Astra and Sol, released the same day.