Microsoft Unveils Cheaper In-House AI Models to Take On OpenAI
Microsoft launched seven in-house AI models at Ignite, saying its new MAI Code 2 nears Claude Opus 5.5 on coding tests at about 5% of the price.
Microsoft unveiled seven new artificial intelligence models built by its own researchers at its Ignite 2026 conference on Tuesday. The pitch is aimed squarely at rivals OpenAI and Anthropic. The company says its new coding model, MAI Code 2, delivers about 96% of the coding performance of Anthropic’s top model, Claude Opus 5.5, at roughly 5% of the price. The performance and price figures come from Microsoft’s own testing and have not been independently verified.
The launch came from Microsoft AI, the company’s consumer and model-building division, which is led by chief executive Mustafa Suleyman. It is Microsoft’s broadest push yet to sell its own models, which it brands MAI, alongside the OpenAI and Anthropic technology it already offers to customers.
What Microsoft announced
According to Microsoft AI’s announcement and coverage by the technology site Thurrott.com, the lineup covers coding, cybersecurity, images, voice and medical transcription:
- MAI Code 2, for everyday enterprise software work, available in GitHub Copilot, Microsoft’s AI coding assistant.
- MAI Code 2 Flash, a lighter model for routine jobs such as code explanations and quick fixes, offered in GitHub Copilot and in a limited preview on Microsoft Foundry, the company’s platform for building AI applications.
- MAI Cyber 2 Flash, built to run around the clock and look for security weaknesses across large codebases.
- MAI Image 2.7, for image generation and editing.
- MAI Realtime 1, Microsoft’s first live speech-to-speech model, priced from $0.01 per audio minute.
- Dragon Copilot Transcribe, a medical transcription product built on the MAI Transcribe 2 model.
- Decision 1.1, a routing model that sends a request to a larger, more expensive model only when needed.
Microsoft says MAI Code 2 reaches 98% of the performance of OpenAI’s GPT-6.1 Sol at about 40% of the price. The scores are an average across four coding benchmarks: DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1. The company also said developers in a blind side-by-side test preferred MAI Code 2’s front-end work over GPT-6 Sol’s 64% of the time. Microsoft added that the model is not yet tuned for harder problems, such as fundamental science research or very long-running agentic tasks.
The price argument
Cost is the center of Microsoft’s pitch. Suleyman said in his Ignite keynote that the company tries to balance the two goals, telling the audience, “We co-optimize quality and price together,” Thurrott.com reported. He added that Microsoft does not “just hill-climb on frontier quality alone.”
Microsoft made similar claims across the rest of the lineup. It said MAI Code 2 Flash matches GPT-6 Luna at 65% of its price and Claude Haiku 5.5 at 20% of its price. MAI Cyber 2 Flash, it said, beats the quality of OpenAI’s GPT-6 Astra on a cybersecurity benchmark called CyberGymE2E at under 2% of the cost per task. The AI news site MLQ.ai reported that the security model is priced at $0.10 per million input tokens and $0.50 per million output tokens. Microsoft also said Decision 1.1 is 50 times faster and 79 times cheaper than GPT-6.1 Sol.
Outside checks are scarce. MLQ.ai noted that Microsoft did not publish independent benchmark results. The only third-party ranking cited is from Artificial Analysis, an AI benchmarking firm, which Microsoft said placed MAI Image 2.7 first for image editing and fourth for text-to-image generation. Microsoft’s materials also contain small inconsistencies in model names, and both Thurrott.com and MLQ.ai flagged them.
Why it matters
Microsoft is one of OpenAI’s biggest backers and sells OpenAI models through its Azure cloud. It also offers Anthropic models through some of its products. By benchmarking its own models directly against both partners and stressing lower prices, Microsoft is positioning itself as a model maker in its own right, not only a reseller.
The company is also tying the models to its own hardware. Thurrott.com reported that all seven run in production on Maia 200, Microsoft’s second-generation in-house AI chip for data centers. Microsoft says the chip delivers 1.7 times more tokens per watt. Running its own models on its own chips could help Microsoft lower serving costs and depend less on outside suppliers.
What comes next
Microsoft said it has started building customized versions of its models for specific customers. The first is a model for analyzing 5G network operations. It also plans a program called MAI Licensed Weights, which would let organizations deploy and adapt Microsoft’s models in their own Azure environments. The company said the program is “coming soon” through its Frontier Company program but gave no date. Several of the new models are available now in GitHub Copilot and Microsoft Foundry, while MAI Realtime 1 is open to customers who register their interest.


