Microsoft AI launches seven MAI models and MAI Licensed Weights

Microsoft AI launches seven MAI models and MAI Licensed Weights

Microsoft AI (MAI) announced "a family of seven new world class models" pitched on price-performance. Its stated reason: token spend is climbing, and customers want models cheap enough to run agents around the clock on real engineering work. The announcement does not state which seven models make up the count. The text describes MAI Cyber 2 Flash, MAI Code 2, MAI Code 2 Flash, MAI Image 2.7, MAI Image 2.7 Flash, MAI Realtime 1, MAI Transcribe 2 (via Dragon Copilot Transcribe) and Decision 1.1. Every comparison below is MAI's own claim.

Security: MAI Cyber 2 Flash is aimed at monitoring and control agents that defend and patch live production systems. It builds on earlier MAI Cyber models. MAI says MAI Cyber accounts for almost 40% of scans in MDASH, and that 90% of scans that started with MAI continued to use it. The new model is priced 86% lower on output tokens than MAI Cyber 1.1 Flash.

Coding: MAI Code 2 targets everyday enterprise coding tasks. Averaged over four benchmarks (DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1), MAI says it delivers 98% of the performance of GPT-6.1 Sol at 40% of the price, and 96% of Claude Opus 5.5's performance at 5% of the price. In a blind, side-by-side evaluation of frontend work, developers preferred MAI Code 2 over GPT-6 Sol 64% of the time. MAI Code 2 Flash is meant for high-volume work such as code explanations, targeted edits and routine fixes. MAI says it performs comparably to GPT-6 Luna and Claude Haiku 5.5 at 65% and 20% of their prices respectively, across three benchmarks (DeepSWE 1.1, Terminal Bench 2.1 and SWE-Atlas). MAI says plainly that both coding models are "explicitly not yet tuned for harder problems, like fundamental science research and very long horizon agentic work."

Images: MAI calls MAI Image 2.7 the best image editing model in the world, citing the #1 spot on Artificial Analysis and a price of one-fifth that of GPT Image 2.6 Sunburst (max). In text-to-image generation it ranks #4, which MAI describes as the second-best lab. MAI Image 2.7 Flash targets high-volume use at about $10 per 1,000 images.

Speech and transcription: MAI Realtime 1 is MAI's first live speech-to-speech model, built for low-latency conversation. MAI says it outperformed GPT-Realtime 2.1 on external and Dynamics 365 benchmarks, and prices it from $0.01 per audio minute. Dragon Copilot Transcribe, built on MAI Transcribe 2, is live in Dragon Copilot. MAI says it powers 100% of doctor-patient conversation transcriptions there, covers 10 million encounters a month, halves error rates in multilingual workflows and is used by a third of US physicians.

Routing: Decision 1.1 is model routing built into applications so that bigger, pricier models are used only when needed. MAI says that on the Artificial Analysis Decision Index Microsoft is ahead of Jev, Google and OpenAI while being faster and cheaper. It is used by 4,000 customers and is described as 50x faster and 79x cheaper than GPT-6.1 Sol.

MAI Licensed Weights: The announcement's headline offer. MAI says leaders told it that no one wants to pour their IQ into someone else's model, and it questions whether open weights can be trusted on provenance or behavior. Under Licensed Weights, a customer takes a frontier MAI model, optionally customizes it with proprietary data, and runs it in its own Azure environment. The customer controls who can access it, and the data and adaptations stay the customer's. MAI will work with the customer through Frontier Company to deliver the custom model and embed it in the customer's products and agents. As an earlier example from MAI's Custom Models offering, AT&T developed an MAI custom model for 5G network operation analytics that, MAI says, outperformed GPT-6.1 Sol at 1/50th of the cost.

Key facts

  • Microsoft AI announced a family of seven MAI models pitched on price-performance, plus MAI Licensed Weights, which lets customers customize a frontier MAI model with their own data and run it in their own Azure environment.
  • MAI says MAI Code 2 delivers 98% of GPT-6.1 Sol's performance at 40% of the price, and 96% of Claude Opus 5.5's performance at 5% of the price, averaged over four coding benchmarks.
  • MAI Image 2.7 is claimed to rank #1 for image editing on Artificial Analysis at one-fifth the price of GPT Image 2.6 Sunburst (max); it ranks #4 in text-to-image generation.
  • MAI Code 2 is in GitHub Copilot now; MAI Code 2 Flash is in Copilot and in limited-access preview in Microsoft Foundry; both image models are in Foundry; MAI Realtime 1 is limited to MAI Labs with interest registration.
  • MAI states that its coding models are not yet tuned for fundamental science research or very long-horizon agentic work, and all benchmark and price comparisons are the vendor's own.

Why it matters

Microsoft's own AI group is putting first-party models up against OpenAI and Anthropic offerings, and it is competing on price rather than raw capability. The central claim is that MAI Code 2 gets close to the performance of GPT-6.1 Sol and Claude Opus 5.5 at 40% and 5% of their prices. The Licensed Weights offer adds a second pitch: a customer can own a customized model and run it inside its own Azure environment, rather than building on someone else's hosted model or on open weights of uncertain provenance.

Who it affects

Enterprise developers and teams running coding agents, who are the stated target of MAI Code 2 and Code 2 Flash and see the models in GitHub Copilot. Security teams and existing MDASH customers, who get MAI Cyber 2 Flash at a lower output-token price. Clinicians in Dragon Copilot, where MAI says Dragon Copilot Transcribe is used by a third of US physicians. Enterprises that want a custom model under their own control, the audience for MAI Licensed Weights, of which AT&T's custom model is the cited example.

How to use it

MAI Code 2 is available in GitHub Copilot today. MAI Code 2 Flash is in GitHub Copilot and in a limited-access preview in Microsoft Foundry. MAI Image 2.7 and MAI Image 2.7 Flash are available in Microsoft Foundry. MAI Realtime 1 can be tried in MAI Labs, with interest registration rather than general availability. Dragon Copilot Transcribe is live in Dragon Copilot. For MAI Licensed Weights, the announcement describes the flow: take a frontier MAI model, optionally customize it with proprietary data, run it in your own Azure environment, and work with MAI through Frontier Company to deliver and embed it. No pricing is given for the coding models in absolute terms, only as percentages of competitors' prices. No pricing, availability date or terms are given for MAI Licensed Weights.

How solid is it

This is Microsoft AI's own announcement, so every ranking and comparison is the vendor's claim. No independent verification of the benchmark results is cited, apart from the stated Artificial Analysis rankings. The price comparisons are relative percentages, not absolute prices per million tokens. The 64% developer preference result comes from a blind, side-by-side evaluation of frontend work against GPT-6 Sol, while the 98% figure compares against GPT-6.1 Sol; the source uses both names and does not explain the difference.

Risks and caveats

MAI says the coding models are explicitly not yet tuned for fundamental science research or very long-horizon agentic work, so the price-performance claim covers everyday enterprise tasks only. The announcement does not state which seven models make up the family. The baseline for the "2% of the cost" figure on frontier cyber defense is not stated. The announcement also names "Jev" as a competitor without explanation, and does not say what Frontier Company is or which model powers Decision 1.1 routing. MAI Realtime 1 is available only through MAI Labs with interest registration, and Code 2 Flash is a limited-access preview in Foundry.

“We’ve heard again and again from leaders that no one wants to pour their IQ into someone else’s model.”

— MAI announcement, Microsoft AI