Microsoft Unveils Coding and Security AI Models

Microsoft Unveils Coding and Security AI Models
Wednesday, November 18, 2026
  • •Microsoft announced coding, security, and decision models at Ignite 2026.
  • •MAI Code 2 delivered 96% of Claude Opus 5.5’s performance at 5% of its cost.
  • •MAI Cyber 2 Flash scored 92.3% on CyberGym-E2E at less than 2% of the cost per task.

Microsoft unveiled coding model MAI Code 2, lightweight MAI Code 2 Flash, and security model MAI Cyber 2 Flash at its annual Ignite 2026 technology event in San Francisco on the 17th local time. It also officially launched Microsoft-Decision-1.1 for AI agent decision-making. Microsoft said it aims to reduce reliance on external models and strengthen its own models.

MAI Code 2 analyzes large code repositories and can modify features across multiple files, fix errors, and run tests and validation. It can also create website and application screens from screenshots or UI mockups. The mixture-of-experts model has 2.8 trillion parameters, with 80 billion activated during inference, supports a context window of up to 1.05 million tokens, and accepts image input. Microsoft pretrained it on about 50 trillion tokens, then used reinforcement learning in the real development environment of GitHub Copilot.

In Microsoft’s evaluations, MAI Code 2 averaged 98% of GPT-6.1 Sol’s performance across four coding benchmarks at 40% of its cost. Compared with Claude Opus 5.5, it reached 96% of the performance at 5% of the cost. A blind test of UI, game, and data visualization outputs included 258 evaluators; 67% chose MAI Code 2’s results over GPT-6 Sol’s. The cost comparison used promotional prices discounted by 50%.

MAI Code 2 Flash, a lightweight model for iterative development work, has 398 billion parameters, of which 12 billion are activated, and supports the same 1.05-million-token context window as the larger model. Microsoft said it offers coding performance close to GPT-6 Luna and Claude Haiku 5.5 at 65% and 20% of their respective costs. AT&T said a version of Flash adapted to its work outperformed GPT-6.1 Sol on 5G network analysis at one-fiftieth the cost. Woven by Toyota said fine-tuning it for automotive coding standards improved code generation for real work while costing less than its previous state-of-the-art model. Both coding models are available in GitHub Copilot, while Flash is also available on a limited basis in Microsoft Foundry.

MAI Cyber 2 Flash specializes in software vulnerability detection and security incident investigations. It scored 92.3% on CyberGym-E2E, an evaluation of vulnerability detection and reproduction, close to GPT-6 Astra’s 91.5%, at less than 2% of the cost per task. Microsoft plans to integrate it into security agent systems for some customers starting in mid-December. Microsoft-Decision-1.1 handles simple decisions such as tool selection and task classification quickly and at low cost; this version adds image input. More than 4,000 customers already use it.

Microsoft AI CEO Mustafa Suleyman told Reuters that the results were not limited to a single model or a temporary outcome, and said the company had built a system that optimizes performance and price together. Microsoft also announced its MAI Licensed Weights strategy, which licenses model weights so businesses can adapt models to their own work. The company said it is targeting the market for customized enterprise models beyond API offerings, while continuing to develop general-purpose frontier models.

Subscribe to the AIB Newsletter

The fastest, most reliable AI news, straight to your inbox