Microsoft has announced a family of new cyber- and code-focused AI models that it says are competitive with frontier rivals at a fraction of the cost.
At its annual Microsoft Ignite conference, the tech giant unveiled MAI Cyber 2 Flash, a new AI model focused on real-world security tasks such as detecting and patching vulnerabilities.
In the benchmark CyberGym-E2E, which tests AI models on their ability to discover cyber vulnerabilities and patch them without breaking real-world functionality, Microsoft said MAI Cyber 2 Flash scored 92.3 per cent.
This puts it ahead of OpenAI’s GPT-6 Astra, which scored 91.55 per cent, at less than 2 per cent of the cost per task.
The new model will be introduced exclusively within Microsoft’s security products including Codename MDASH, a new agentic code scanner within Microsoft Defender. Microsoft said the multi-model system will work alongside security and engineering teams to identify and remediate code vulnerabilities.
Mustafa Suleyman, chief executive of Microsoft AI, said: “The last few months have been a warning shot for everyone. It's very clear that attackers are gaining momentum, as exploits become cheaper and easier than ever.
“Optimising the price-performance frontier is critical if we're to enable developers to secure their code, systems, and networks against vulnerabilities.”
Alongside the cyber model, Microsoft announced MAI Code 2, its new flagship code generation model intended for production code, and MAI Code 2 Flash, a lightweight model that Microsoft said is optimised for high-volume tasks such as code explanation, routine fixes or rapid iteration of code.
In DeepSWE 1.1, a benchmark that tests how well AI models can handle complex coding problems, Microsoft said MAI Code 2 scored 73.2 per cent versus GPT-6.1 Sol’s 74.6 per cent and Claude Opus 5.5’s 69 per cent. Microsoft added the model achieved this result with an average cost per task of $0.309, compared to Sol’s $0.761 and Opus’ $6.306.
Across four common coding benchmarks, DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas and FrontierCode v1.1, Microsoft said its model delivered 96 per cent of Claude Opus 5.5’s performance at 5 per cent of the cost.
MAI Code 2 Flash is intended as an even more economical coding model, with an average cost per task across the above benchmarks of $0.052, compared with $0.080 for OpenAI’s cheapest model GPT-6 Luna and $0.319 for Anthropic’s lightweight Claude Haiku 5.5.









Recent Stories