Microsoft Expands Its In-House AI Model Lineup at Ignite, Pricing Against OpenAI and Anthropic

New MAI models for coding, cybersecurity, images, voice and transcription arrive five months after the company’s first seven-model launch. Microsoft’s pitch centers on cost per task.

Microsoft AI announced a second wave of in-house models on November 17 at its Ignite conference in San Francisco. The release adds models for coding, cybersecurity, image editing, real-time voice and transcription, along with a routing model that decides which model should handle a given request. Microsoft says each one performs close to, or ahead of, comparable models from OpenAI and Anthropic at a lower price.

The announcement follows the company’s Build conference on June 2, where Microsoft AI CEO Mustafa Suleyman introduced the first seven models in the MAI family. Several of those have now been replaced by newer versions.

For enterprise marketing leaders, most of whom reach these models through Microsoft 365 Copilot, Dynamics 365, GitHub or Microsoft Foundry, the practical news is that Microsoft’s own models are taking on more of the work inside products they already license. The image, voice and routing models are the ones most likely to touch marketing and customer experience operations.

What Microsoft announced

Microsoft’s own post describes seven new models. Axios, which interviewed Suleyman, counted six. The difference comes down to what is counted as a model, and the lineup is as follows.

MAI Code 2 and MAI Code 2 Flash. Microsoft says MAI Code 2 delivers 98% of the coding performance of OpenAI’s GPT-6.1 Sol at about 40% of the price, and 96% of Anthropic’s Claude Opus 5.5 at about 5% of the price, averaged across four coding benchmarks. The smaller Flash version is positioned against GPT-6 Luna and Claude Haiku 5.5. Both are generally available in GitHub Copilot, and Flash is in private preview in Foundry. Microsoft notes that MAI Code 2 is not yet tuned for scientific research or very long-running agent tasks.

MAI Image 2.7 and MAI Image 2.7 Flash. Microsoft says MAI Image 2.7 ranks first for image editing on the Artificial Analysis leaderboard at one fifth the price of OpenAI’s top image model, and fourth for text-to-image generation. The Flash version is priced at about $10 per 1,000 images. Microsoft’s Ignite Book of News lists both as in preview.

MAI Realtime 1. This is Microsoft’s first live speech-to-speech model, built for low-latency voice conversations. Microsoft says it outperformed OpenAI’s GPT-Realtime 2.1 on external benchmarks and on Dynamics 365 benchmarks, with pricing starting at $0.01 per audio minute. It is available to try in MAI Labs.

MAI Transcribe 2. The new transcription model ships first inside Dragon Copilot, Microsoft’s clinical documentation product. Microsoft says it handles about 10 million patient encounters a month and has cut multilingual error rates in half.

MAI Cyber 2 Flash. This is a new category for the MAI family. The model finds, validates and fixes code vulnerabilities, and Microsoft says it at least matches OpenAI’s GPT-6 Astra on a cybersecurity benchmark at under 2% of the cost per task. It is in preview only inside Codename MDASH, Microsoft’s agent-based code security system, for eligible organizations.

Decision 1.1. This model routes each request to the smallest model that can handle it. Microsoft says it is 50 times faster and 79 times cheaper than using GPT-6.1 Sol for the same routing decisions, and that 4,000 customers use it. It routes requests in Copilot and is generally available in Foundry.

Microsoft also said its second-generation Maia 200 chip is now running these models in production.

From Build to Ignite

The June launch included MAI Thinking 1, a reasoning model, plus models for coding, images, transcription and voice. At the time, GeekWire reported that Suleyman framed the effort as a matter of long-term self-sufficiency for Microsoft. He also stressed that the models were trained from scratch on licensed data, without distilling the output of other companies’ models.

Microsoft has built much of its AI business on OpenAI’s models and has invested about $13 billion in that company, according to GeekWire. It has also committed up to $5 billion to Anthropic. Both companies’ models remain available in Foundry, which Microsoft says now offers more than 11,000 models.

The Ignite release does not include a new general-purpose reasoning model. Axios reported that Suleyman said Microsoft is making significant progress toward its own frontier general-purpose model. Other outlets summarizing that interview said he gave no timeline.

In his Ignite keynote, Suleyman put the emphasis on price and performance together. “We co-optimize quality and price together,” he said, according to Thurrott.com.

Custom models and licensed weights

Two items in the announcement concern how enterprises can adapt the models.

Microsoft said AT&T built a custom MAI model for analyzing 5G network operations and that it outperformed GPT-6.1 Sol at 50 times lower cost. Microsoft has not published details of that evaluation.

The company also introduced MAI Licensed Weights, which would let an organization take an MAI model, tune it on its own data and run it in its own Azure environment. Microsoft says the customer keeps control of its data and of the adapted model. Thurrott.com reported that the offering is coming soon through Microsoft’s Frontier Company program, with no date given.

Separately, Foundry added preview features for training supported models on organizational data.

What marketing and CX leaders should weigh

The benchmarks are Microsoft’s. The performance and price comparisons come from Microsoft’s own testing on benchmarks it selected. The Artificial Analysis image ranking is a third-party leaderboard, but the cost comparisons are Microsoft’s math. Teams considering a switch for a specific workload will want to test on their own content and tasks.

Availability varies widely. The coding models are generally available. The image models are in preview, the voice model is in a labs environment, and the cybersecurity model is limited to one preview program. Marketing teams interested in MAI Image 2.7 for production creative work should confirm status and commercial terms before planning around it.

Routing changes which model does the work. Decision 1.1 chooses a model per request inside Copilot. That can lower cost, and it also means the model behind a given answer may vary. Organizations with brand, compliance or quality requirements tied to a particular model may want to ask how routing can be configured or audited.

Voice is the closest fit for customer experience. A lower-cost real-time voice model tested against Dynamics 365 workloads points toward contact center use. Microsoft has not said when MAI Realtime 1 will be available in Dynamics 365 Contact Center or under what terms.

Data lineage is part of the pitch. Microsoft has said since June that the MAI models are trained on licensed data without distillation from other labs. For brand and legal teams that have asked vendors about training data for generative image tools, that is a claim to ask Microsoft to document.

Custom models raise ownership questions. Licensed Weights and custom models put more control with the customer, and they also add work. Someone has to own the training data, the evaluation and the ongoing upkeep of a tuned model.

Microsoft’s post describes security, alignment and containment as priorities for its agent work in 2027. It gives no roadmap for further model releases.

Microsoft Customer Insights
The Agile Brand Guide®
Privacy Overview

This website uses cookies so that we can provide you with the best user experience possible. Cookie information is stored in your browser and performs functions such as recognising you when you return to our website and helping our team to understand which sections of the website you find most interesting and useful.