Microsoft pushes its own AI models to reduce costs

Microsoft has introduced seven in-house AI models for code, security, image, voice, and routing. The goal is clear: to lower the cost of AI agents while keeping these models within its products and Azure.

Microsoft unveiled seven specialized AI models at Ignite, designed to replace external models in its own products and enterprise clients. The group is betting on in-house alternatives for code, cybersecurity, image, voice, and decision-making, with a central argument: reducing the cost of continuously running AI agents. At this stage, the strategy primarily targets OpenAI and Anthropic, but it also fits into the broader relationship Microsoft maintains with its partners and Azure.

Des modèles maison pour reprendre la main sur la facture

The core message is economic. Microsoft explains that the use of always-active AI agents increases token consumption, these text units used to bill and run models. Hence the arrival of MAI Code 2, MAI Code 2 Flash, MAI Cyber 2 Flash, MAI Image 2.7, MAI Realtime 1, Decision 1.1, and Microsoft-Retriever-1. The group presents them not as lab curiosities but as building blocks ready to be integrated into GitHub Copilot, Microsoft 365, Fabric, Foundry, MDASH, or Dynamics 365.

On paper, the promise is value for money. Microsoft claims that MAI Code 2 reaches 98% of GPT-6.1 Sol's performance on four internal benchmarks, at 40% of the cost. Against Claude Opus 5.5, the score would drop to 96% of performance for 5% of the cost. The company also specifies that, in blind comparisons on interfaces, games, or visualizations, human evaluators preferred its outputs 64% of the time. The model is priced at $0.50 per million input tokens and $2.50 per million output tokens during the promotional period, until May 17, 2027.

Code, cybersécurité et routage : les usages les plus sensibles

Code and security are the two most strategic areas. MAI Code 2 Flash targets explanations, targeted edits, and common corrections, with an even more aggressive pricing. MAI Cyber 2 Flash is designed for vulnerability testing and security scans. Microsoft states it solves 92.3% of tasks on the CyberGym-E2E benchmark, compared to 91.5% for GPT-6 Astra, while using less reasoning budget. On this test, the cost per task would drop to $0.45, versus $34.38 for GPT-6 Astra. A saving that is anything but trivial.

The company also introduces Decision 1.1, a routing and classification model based on Google's Gemma 4 26B A4B. This model chooses developer-approved actions, returns calibrated confidence levels, and can abstain when information is lacking. In short, it helps prevent an application from acting too quickly when uncertain. On Artificial Analysis's Decision Index v2.1, it scores 68 points, six more than its predecessor. Microsoft also says it processes requests 2.2 times faster and 4.4 times cheaper than a traditional large language model in certain Copilot and Cowork uses.

L’image et la voix entrent dans la même logique industrielle

MAI Image 2.7 occupies another key segment: image editing. Microsoft presents it as the top model in the Artificial Analysis ranking for image editing, with 1,215 Elo points and a cost significantly lower than GPT Image 2.6 Sunburst. The model can modify the color, material, or position of an object without affecting the rest of the image, and generate transparent backgrounds. Its Flash version powers Canvas, a new image editor in MAI Labs. For Microsoft, this type of tool must be good enough for creative teams but affordable enough to be used without counting every attempt.

On the voice side, MAI Realtime 1 is Microsoft's first in-house speech-to-speech model. It understands interruptions, waits for a speaker to finish their sentence, and even works with background noise. Compatible with the OpenAI Realtime API, it can theoretically be adopted without massive application redesign. Microsoft is already testing it with Agoda and Reservations.ai, and plans to integrate it into Dynamics 365 Contact Center as well as Dragon Copilot. In the latter, Dragon Copilot Transcribe, based on MAI Transcribe 2, already handles 10 million doctor-patient conversations per month.

Ce que Microsoft garde en interne, et ce qu’il ouvre aux entreprises

The launch is not limited to visible products. With MAI Licensed Weights, Microsoft offers certain companies the ability to fine-tune MAI models with their own data in their Azure environment. Data and customizations remain with the client, and Microsoft states it only uses this data for its own models with explicit consent. The offer is designed as an alternative to open weights, whose origin and behavior can sometimes be difficult for companies to verify.

In practice, this component could influence purchasing decisions of large organizations wanting to keep control over their models without starting from scratch. Microsoft notably cites AT&T, where a customized model for 5G network operations analysis outperformed GPT-6.1 Sol on standard tasks, at about one-fiftieth of the cost. The availability and price of MAI Licensed Weights remain to be confirmed [to be verified]. Meanwhile, the group continues to deploy its models at different paces: some are already available in GitHub Copilot, Foundry, or Vercel, while MAI Cyber 2 Flash remains reserved for Microsoft security products.

Ultimately, Microsoft seeks to reduce its visible dependence on partners while maintaining access to cutting-edge models when needed. It is not exactly a break, but rather an ongoing negotiation with the bill. And in AI, the bill always ends up speaking.

Key points

  • Microsoft presented seven models at Ignite.
  • MAI Code 2 targets GitHub Copilot.
  • MAI Cyber 2 Flash achieves 92.3% on CyberGym-E2E.
  • Decision 1.1 is based on Gemma 4 26B A4B.
  • MAI Realtime 1 is compatible with the OpenAI Realtime API.
  • Dragon Copilot handles 10 million monthly conversations.

By the numbers

  • 7 models — announced by Microsoft at Ignite.
  • 1 million tokens — context window of MAI Cyber 2 Flash.
  • 92.3% — MAI Cyber 2 Flash score on CyberGym-E2E.
  • 1,215 Elo points — MAI Image 2.7 rating on Artificial Analysis.
  • 10 million conversations per month — transcriptions handled by Dragon Copilot.

To read