Microsoft launches seven MAI models: what changes for businesses
Microsoft presents seven MAI models at Ignite 2026 for coding, security, images and voice: what the company claims and what changes for businesses.

At Ignite 2026, Microsoft presented a family of seven MAI models, developed by its Microsoft AI division, for programming, cybersecurity, images and voice. According to the company, the new MAI models come close to the performance of those from OpenAI and Anthropic at lower costs.12 The figures circulating, however, are results from tests chosen by Microsoft, with different conditions for each model, and some tools are available only to a restricted group of customers.12
What are the seven MAI models
For programming there are MAI Code 2 and the lighter variant Code 2 Flash. Image 2.7 and Realtime 1 cover images and real-time voice, while Cyber 2 Flash is designed for cyber defense.1 The package also includes Decision 1.1, now generally available, and Transcribe 2, which already supports transcription in Dragon Copilot. The offering is complemented by a model that routes requests to the most suitable one.1 The count of seven depends on how the products are counted. Axios speaks of six new models, because it considers only the novelties; Microsoft, instead, adds the Decision update and a transcription system that was already in operation.1
What Microsoft claims about the comparisons
For MAI Code 2 the company indicates a quality equal to 98% of that of GPT-6.1 Sol at 40% of the price, as an average of four benchmarks, meaning standardized tests, on everyday enterprise coding. Compared with Claude Opus 5.5, the figure is 96% of the quality at 5% of the cost.1 In these calculations, the price of MAI Code 2 is discounted by 50%. For Code 2 Flash Microsoft claims results comparable to GPT-6 Luna while spending 65%, and to Claude Haiku 5.5 while spending 20%, on three routine coding benchmarks, again with a promotional price.1 The technical documentation warns that MAI Code 2 is not yet optimized for the most demanding tasks, such as basic scientific research or agents tasked with planning over long time spans.1 Cyber 2 Flash achieves 92.3% success on CyberGym-E2E, a test on discovering software vulnerabilities, against 91.5% for GPT-6 Astra. For this model Microsoft claims less than 2% of the cost per task and half the reasoning budget of the competitor.1 For Image 2.7 Microsoft indicates first place in image editing in the Artificial Analysis leaderboard as of November 16, 2026, with 1,215 Elo points, a 4-point lead and 5,478 votes. The cost would be about one fifth of that of GPT Image 2.6 Sunburst (max).1
Who can use them today
Availability varies from model to model, ranging from preview, general release and integrations already operational. Some models are in Microsoft Foundry or integrated into Copilot, while the company indicates that third-party offerings remain available in Foundry.1 The cybersecurity model will be offered to a limited group of customers. Microsoft links it to the rise in software vulnerabilities that AI tools increasingly bring to light.2
Suleyman’s strategy
Mustafa Suleyman, head of Microsoft AI, told Reuters that building technology specific to individual domains helps contain costs while matching rivals and, in some cases, surpassing them on benchmarks. He described the work as a system that optimizes performance and price together.2 According to Axios, Suleyman presents the logic of the models as an attempt to fit more tokens, knowledge and experience into the same number of parameters.1 Suleyman also said that alignment, meaning research to prevent AI from taking harmful actions and to make it behave as intended, is his priority. He cited as a driver the recent cyberattack on the Hugging Face platform attributed to OpenAI agents.2
What changes
The cost figures, however, should be read with caution, because they rest on discounted prices, on benchmarks selected by the company and on different configurations for each model. For a business, what will matter most is the result obtained on its own code, with its own tools and its own review rules.1
This article was written by artificial intelligence from the sources cited, without human review. Automatic newsroom by Redazia.


