Seven new models at scalable price-performance
- Source
- microsoft.ai
- Author
- gmays
- Date

Context
At its Ignite conference, Microsoft's in-house model group, Microsoft AI (MAI), announced seven models built for separate jobs and sold mainly on cost. Most come with a claim that they match or beat a named model from OpenAI or Anthropic for a fraction of the price. The pitch is aimed at companies running AI agents around the clock, whose bills grow with every processed.
The coding pair shows how the pitch works. Microsoft says MAI Code 2, averaged over four coding benchmarks, reaches 98% of GPT-6.1 Sol's score at about 40% of its price and 96% of Claude Opus 5.5's at about 5%. It also says the model is not yet tuned for fundamental science research or very long agentic tasks. Code 2 and the cheaper Code 2 Flash, aimed at high-volume routine work, are available in GitHub Copilot, and Flash is in limited preview in Microsoft Foundry, Microsoft's cloud platform for building on AI models.
The rest of the family covers security, images, speech and routing. MAI Cyber 2 Flash is meant to run defensive security agents continuously. MAI Image 2.7 and its Flash variant generate and edit images. MAI Realtime 1 is Microsoft's first model that listens and replies in speech directly. Dragon Copilot Transcribe, built on MAI Transcribe 2, transcribes doctor and patient conversations. Decision 1.1 is a router: it sits in front of other models and sends each request to a large, expensive model only when the task needs one.
Most of these comparisons are Microsoft's own measurements. The article names the benchmarks behind some of them but does not explain how the tests were run. Two rankings are credited to Artificial Analysis, an independent benchmarking firm: Decision 1.1's lead on its Decision Index, and MAI Image 2.7's first place in image editing. Artificial Analysis announced the image editing result itself, placing MAI-Image-2.7 first on its image editing leaderboard at about one sixth the price of the four models behind it. The other price-performance claims remain Microsoft's own until outside tests confirm them.
- token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
Checking sign-in…
Loading comments…





