Skip to content

Microsoft: Smaller, Cheaper AI Could Win

Microsoft unveiled specialized AI models for coding, cybersecurity and image generation at its Ignite conference. It says MAI Code 2 delivers 96% of Claude Opus 5.5’s performance on common enterprise coding tasks at 5% of the cost, but that cost uses a 50% promotional discount. Microsoft ran three of the four performance evaluations on its own platform. Code 2 is available in GitHub Copilot, though Microsoft says it is not yet optimized for fundamental scientific research or very long autonomous tasks.

In Microsoft’s evaluations, MAI Cyber 2 Flash completed 92.3% of the assessed vulnerability-discovery tasks, against 91.5% for GPT-6 Astra; Microsoft excluded 92 of 905 tasks because of overlap with training data. It has not published a direct comparison with OpenAI’s restricted GPT-6 Cyber model. Cyber 2 Flash is scheduled to begin replacing an older model for eligible MDASH customers in mid-December.

MAI Image 2.7 leads an independent image-editing ranking by four points, but its uncertainty range overlaps that of the runner-up; it ranks fourth on the separate text-to-image list.

Source

Share