Microsoft Launches Seven AI Models and Weight Control

Microsoft launches seven new AI models aimed at code, images, voice, and automated decision-making. The company also introduces MAI Licensed Weights, an offering that allows customers to customize models and run them in their Azure environment. The claimed advantages over competing models are based on comparisons provided by Microsoft.
Companies Can Control Their MAI Models
AT&T has already developed a MAI model tailored for analyzing its 5G network operations. Microsoft claims that this model outperforms GPT-6.1 Sol for 1/50th of the cost. The company also reports that AT&T has observed Code 2 performance in complex scenarios comparable to the most advanced models, at a significantly lower price than other offerings.
This customization, already offered by Microsoft, is extended with MAI Licensed Weights. Customers can adapt a model to their proprietary data and then run it in their own Azure environment. Microsoft assures that customers control access and retain ownership of their data as well as their adaptations. Support from Frontier Company is planned for the delivery of the customized model and its integration into the customer's products or agents.
Medical Transcription and Voice Conversations
Deployed within Dragon Copilot, Dragon Copilot Transcribe is based on MAI Transcribe 2. Microsoft states that it ensures 100% of transcriptions for exchanges between doctors and patients in this tool, for 10 million consultations monthly. According to the group, it is used by one-third of American doctors.
Microsoft claims a halving of error rates for multilingual usage. It also presents MAI Transcribe 2 as the fastest and most accurate multilingual transcription model in the world.
For voice conversations, Microsoft introduces MAI Realtime 1 as its first model capable of transitioning directly from speech to speech with low latency. The group claims it surpasses GPT-Realtime 2.1 in its evaluations, conducted on external tests and Dynamics 365. Pricing starts at $0.01 for one minute of audio. A trial is offered in MAI Labs, which also allows users to express their interest.
Two Models for Creating and Modifying Images
In Microsoft Foundry, users can access MAI Image 2.7 and its Flash variant. The latter targets use cases requiring numerous images, priced at around $10 for 1,000 images.
Microsoft indicates that MAI Image 2.7 ranks first in the image editing category of Artificial Analysis and outperforms GPT Image 2.6 Sunburst (max) for one-fifth of the price. For images created from text, the group announces a fourth-place ranking for the model and positions itself second among laboratories.
Le brief IA que les pros lisent chaque soir
Les 7 actus IA du jour, décryptées en 5 min. Gratuit.
Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.
Choisissez votre rythme
Gratuit · Pas de spam · Désabonnement en 1 clic
Models for Code and Cybersecurity
The programming models primarily target daily business tasks rather than fundamental scientific research or extended agent missions, for which Microsoft acknowledges it has not yet optimized. For MAI Code 2, the group announces 98% of the performance of GPT-6.1 Sol at about 40% of the price, and 96% of that of Claude Opus 5.5 at 5% of the price. These results correspond to an average across DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas, and FrontierCode v1.1, selected for tasks observed in GitHub Copilot.
MAI Code 2, available in GitHub Copilot, has also been adjusted for dashboards and visualizations. Microsoft reports that in a blind comparison of interfaces produced by the models, developers preferred it to GPT-6 Sol in 64% of cases. Its Flash variant targets explanations, targeted edits, routine corrections, and rapid iterations in large volumes. On DeepSWE 1.1, Terminal Bench 2.1, and SWE-Atlas, Microsoft claims comparable performance to GPT-6 Luna for 65% of the price, and to Claude Haiku 5.5 for 20% of the price; this variant is offered in GitHub Copilot and in limited access beta in Microsoft Foundry.
In cybersecurity, MAI Cyber 2 Flash is intended for agents who continuously monitor, defend, and correct production systems. Microsoft claims it outperforms GPT-6 Astra on CyberGym-E2E with less than 2% of the cost per task. The group states it relies on decades of proprietary data and security expertise, facing attackers who, according to them, have abundant and inexpensive offensive tools. It prioritizes security, alignment, and containment for agents in 2027 and beyond.
This new version extends a range already employed by defenders: according to Microsoft, MAI Cyber accounts for nearly 40% of MDASH analyses, and 90% of analyses engaged with MAI have continued to use it. Tokens generated by MAI Cyber 2 Flash cost 86% less than those of MAI Cyber 1.1 Flash. Microsoft presents the new version to existing MDASH clients as a more efficient and cost-effective offering.
Model Routing and Maia 200 Accelerator
Decision 1.1, the new version of Microsoft's decision model, provides applications with routing between models. This mechanism aims to only engage the largest and most expensive models when their use is necessary. Microsoft claims to have 4,000 clients and to outpace Jev, Google, and OpenAI in the Artificial Analysis Decision Index, with a faster and cheaper offering. Against GPT-6.1 Sol, the group announces a speed 50 times greater and a cost 79 times lower.
On the hardware side, its second-generation AI accelerator Maia 200 is now in production and runs its largest models. The group claims a yield of 1.7 times more tokens per watt, without specifying the point of comparison.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.