← News

A Microsoft AI presenter on stage at Ignite 2026 in front of the new MAI model lineup: Code, Decision, Cyber, Image, Transcribe, Custom Models and Licensed Weights

17 Nov 2026

Microsoft AI

Microsoft launches seven cheaper MAI models to undercut OpenAI and Anthropic

Microsoft AI released MAI Code 2, Code 2 Flash, Cyber 2 Flash, Image 2.7, Realtime 1, Transcribe 2 and Decision 1.1. The pitch is price: near-frontier quality for agents that run around the clock.

Microsoft is still OpenAI's biggest partner and now its loudest discount competitor. Every number here is Microsoft grading its own homework, so watch independent benchmarks.

At Ignite on Tuesday, Nov. 17, Microsoft AI, the company's in-house model group, which it shortens to MAI, released a family of seven models. The pitch on its own page is price against quality: models cheap enough to leave running all day and all night. A model, here, is the program that does the work. Every quality number and every price comparison below is Microsoft grading its own models. Axios, covering the same announcement, counted six specialized models. Microsoft's page lists seven, and the seven names in the dek are the ones on that page.

MAI Code 2 is the coding model, available in GitHub Copilot now. Microsoft says that, averaged over four coding benchmarks, it reaches 98 percent of GPT-6.1 Sol's coding quality at about 40 percent of the price, and 96 percent of Claude Opus 5.5 at about 5 percent of the price. A benchmark is a fixed test used to compare models. The four tests Microsoft names are DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas, and FrontierCode v1.1. Microsoft also says the model is explicitly not yet tuned for harder problems, including fundamental science and agent jobs that run for a very long stretch of steps. A footnote on the Code 2 post says the MAI Code 2 cost in that comparison is shown at a 50 percent promotional price. Axios notes that some of the pricing gaps are promotional discounts. Those lines are Microsoft's, with the discount caution from Axios.

MAI Code 2 Flash is the smaller one, for high-volume routine coding: explanations, small edits, ordinary fixes, another pass. Microsoft compares it with GPT-6 Luna at 65 percent of the price and with Claude Haiku 5.5 at 20 percent. It is in GitHub Copilot, and in a limited preview in Microsoft Foundry, the catalog where companies pick and run models. Those lines are Microsoft's.

MAI Cyber 2 Flash is the security model. Microsoft's page says it matches GPT-6 Astra on a test called CyberGymE2E at under 2 percent of the cost per task. Output tokens, the text the model writes back, are priced 86 percent lower than on the previous MAI Cyber 1.1 Flash. Axios's verb is stronger: it says the model beats GPT-6 Astra on a security benchmark. The "matches" line is the one on microsoft.ai. Both versions are Microsoft talking about its own test.

MAI Image 2.7. Microsoft says it ranks first on Artificial Analysis for image editing, at one fifth the price of GPT Image 2.6 Sunburst at its max setting. Artificial Analysis is an outside scoreboard. The ranking is still Microsoft reporting that scoreboard. A faster version, Image 2.7 Flash, runs about $10 per 1,000 images. Both are in Microsoft Foundry. Those lines are Microsoft's.

MAI Realtime 1 is Microsoft's first live speech-to-speech model. It hears someone talk and talks back, without a separate full transcript step in between. The starting price Microsoft lists is $0.01 per minute of audio.

MAI Transcribe 2 writes speech down as text. Microsoft says it powers Dragon Copilot Transcribe, which handles 10 million doctor-patient encounters a month and cuts error rates in half on multilingual visits. An encounter, here, is one visit's conversation. That volume and that error claim are Microsoft's.

Decision 1.1 is a router. It sends a job to a bigger, pricier model only when the smaller one is not enough. Microsoft says it is 50 times faster and 79 times cheaper than GPT-Sol 6.1, and that 4,000 customers use it. Those figures are Microsoft's.

The same post adds MAI Licensed Weights. A customer can take an MAI model, customize it on the customer's own data, and run it inside the customer's own Azure environment. Azure is Microsoft's cloud. Weights are the learned numbers inside a model. Microsoft's line is that the adapted model and the data stay the customer's.

Microsoft also said its Maia 200 chip, an AI processor it builds itself, is in production and running the company's largest models. That line sits on the same Ignite post. It is a chip note next to the model list, not one of the seven.

The picture is a presenter on the Ignite stage, in a light jacket and yellow shoes, in front of a black screen headed "Frontier Intelligence." The screen shows seven tiles: Code, Decision, Cyber, Image, Transcribe, Custom Models, and Licensed Weights. The source does not name the person on stage. It is Microsoft AI's event photograph.

In plain terms, Microsoft, still OpenAI's biggest partner, put seven of its own models on stage and priced them as the discount option. The percentages are Microsoft's grades, and the coding comparison uses a promotional price. Axios counted six models and used a stronger verb on the cyber test than Microsoft's own page does.

RELATED

ONLINE…

Comments

guidelines

Loading…

Loading…

Sources