Microsoft has introduced a new family of AI models for coding, cybersecurity, image generation, audio and decision-making, positioning them as lower-cost options for businesses running AI applications and agents at scale. The company unveiled the models at Microsoft Ignite 2026, which began on Nov. 17.
Rather than relying on one general-purpose model for every task, the MAI lineup is made up of specialized models. Microsoft said they delivered competitive results against models from OpenAI and Anthropic in its cited evaluations, while costing less to run. Its announcement also pointed to a growing focus on the cost and speed of completing AI tasks, alongside model capability.
For cybersecurity, MAI Cyber 2 Flash is designed to monitor systems, detect threats and help address vulnerabilities. Microsoft said it matched GPT-6 Astra on the CyberGym-E2E test at less than 2% of the comparison model’s cost, and that its output-token costs are 86% lower than those of the previous MAI Cyber 1.1 Flash. The company said MAI models account for about 40% of scans on the MDASH platform, and that 90% of scans started with an MAI model continued using one.
Microsoft also introduced MAI Code 2 and MAI Code 2 Flash for writing, reviewing, editing and debugging software. According to the company, Code 2 delivered about 98% of GPT-6.1 Sol’s performance on everyday coding tasks at 40% of its cost, and 96% of Claude Opus 5.5’s performance at 5% of its cost, based on four benchmarks. Microsoft said developers preferred its output to GPT-6 Sol in 64% of its side-by-side front-end development evaluations. The company cautioned that Code 2 is not yet tuned for more complex work such as basic scientific research or long-running, multi-step agent tasks.
Code 2 Flash is aimed at frequent, routine jobs such as explaining code, fixing common bugs and making targeted edits. Microsoft said it performed close to GPT-6 Luna at 65% of the cost and Claude Haiku 5.5 at 20% of the cost across three everyday coding tests. Code 2 is available through GitHub Copilot, while Code 2 Flash is available there and in limited preview through Microsoft Foundry.
For image work, Microsoft announced MAI Image 2.7 for generating and editing images, and said it ranked first on the Artificial Analysis image-editing benchmark at one-fifth the cost of GPT Image 2.6 Sunburst (max). The company placed it fourth in text-to-image generation on the benchmark. It also described an Image 2.7 Flash version for high-volume generation, priced at about $10 per 1,000 images. The image models are available through Microsoft Foundry.
MAI Realtime 1, Microsoft’s first speech-to-speech model for live conversations, is intended for voice assistants, customer service and other real-time applications. Microsoft said it outperformed GPT-Realtime 2.1 in its evaluations, including tests connected to Dynamics 365, and set pricing starting at $0.01 per minute of audio. The model is available to try through MAI Labs, where businesses can register their interest.
In healthcare, MAI Transcribe 2 powers transcription in Dragon Copilot, Microsoft’s tool for documenting doctor-patient conversations. The company said it supports about 10 million patient-provider interactions each month and has cut multilingual transcription error rates in half. Microsoft also said about one-third of doctors in the United States use Dragon Copilot.
The final model in the announced lineup, Decision 1.1, is designed to route tasks to suitable models so applications can use less expensive options for simpler work. Microsoft said it is about 50 times faster and 79 times cheaper than GPT-6.1 Sol, and is used by about 4,000 customers. The company also cited its lead in speed and cost on the Artificial Analysis Decision Index.
Alongside the models, Microsoft announced MAI Licensed Weights, which lets businesses customize MAI models with their own data and run them in their Azure environments. The company said customers retain control of their data, model changes and access permissions. Microsoft cited its earlier work with AT&T on a custom model for analyzing and operating 5G networks, which it said outperformed GPT-6.1 Sol at about 50 times lower cost.
Source: Jawlah


