Key Takeaways

  • Microsoft has introduced seven MAI models spanning cybersecurity, coding, image editing, voice, transcription, and model routing.
  • The central sales pitch is price-performance, although most benchmark and cost comparisons remain Microsoft-reported.
  • GitHub Copilot, Microsoft Foundry, Azure, and licensed model weights give Microsoft several routes for driving enterprise adoption.

Microsoft is expanding its in-house AI portfolio with seven MAI models aimed at enterprise workloads where inference cost, latency, and operational scale can matter as much as peak benchmark performance.

The Microsoft AI announcement covers MAI Cyber 2 Flash, MAI Code 2, MAI Code 2 Flash, MAI Image 2.7, MAI Realtime 1, MAI Transcribe 2 through Dragon Copilot Transcribe, and Decision 1.1. Together, they span several of the workloads Microsoft already serves through GitHub, Azure, Dynamics 365, and its healthcare products.

Price is the common thread. Microsoft says MAI Code 2 reaches 98% of GPT-6.1 Sol’s coding performance at about 40% of the price. Against Opus 5.5, Microsoft reports 96% of the performance at 5% of the price. MAI Code 2 Flash, designed for frequent routine tasks, reportedly matches GPT-6 Luna at 65% of the price and Haiku 5.5 at 20%.

Those are striking comparisons, but enterprises should read them as vendor claims rather than settled rankings. Deyron Labs notes that the benchmark and pricing comparisons remain Microsoft-reported. Differences in task selection, reasoning budgets, usage tiers, and production behavior can materially affect the economics.

Still, Microsoft’s workload choices are revealing. The narrower focus indicates a shift toward practical capabilities such as targeted code edits, routine fixes, and rapid iterations rather than solely focusing on long-running autonomous tasks. MAI Code 2 and MAI Code 2 Flash are available through GitHub Copilot, with the Flash model also entering limited preview in Microsoft Foundry. The Diary of AI similarly frames the release around Microsoft’s benchmarks, pricing, and Ignite model lineup.

Cybersecurity follows the same pattern. Microsoft reports that MAI Cyber 2 Flash beats GPT-6 Astra's quality on the CyberGymE2E benchmark at less than 2% of the cost per task. MAI Cyber already accounts for almost 40% of scans in MDASH, according to Microsoft, which emphasizes that organizations must run monitoring and control agents to defend and patch live production systems 24-7.

Elsewhere, MAI Image 2.7 reportedly ranks first for image editing on Artificial Analysis and costs one-fifth as much as GPT Image 2.6 Sunburst at its maximum setting.

Voice and healthcare add another angle. MAI Realtime 1 starts at $0.01 per audio minute and outperforms GPT-Realtime 2.1. In healthcare settings, Dragon Copilot Transcribe, built on MAI Transcribe 2, handles 10 million patient encounters per month and has reportedly halved multilingual error rates.

Then there is Decision 1.1. The latest decision model from Microsoft is described as 50 times faster and 79 times cheaper than GPT-Sol 6.1. Its role is practical: direct simpler requests to less expensive models and reserve larger systems for harder computational work.

Microsoft is assembling a vertically integrated AI stack. Maia 200, Microsoft's second-generation AI accelerator, is now in production running its largest models and reportedly delivers 1.7 times more tokens per watt. Microsoft Foundry distributes the models, GitHub Copilot brings them to developers, and MAI Licensed Weights lets customers customize a model and operate it inside their own Azure environment. The benchmarks will require broader validation, but the commercial strategy is clear: make specialized AI inexpensive enough to run continuously throughout daily enterprise operations.