Microsoft Unveils New AI Models for Code and Cybersecurity

Microsoft introduced MAI Code 2 and Code 2 Flash
On November 17, Microsoft announced MAI Code 2 and its lightweight version, Code 2 Flash. The new models are designed to work with program code, including bug fixing, file modification, and performing tasks in repositories.
According to the company, MAI Code 2 achieved an average of 96% of Claude Opus 5.5's results at estimated costs that are 5% of the competitor's usage cost. Compared to GPT-6.1 Sol, the model showed 98% of the result at 40% of the cost.
Microsoft obtained these metrics based on four tests: DeepSWE 1.1, Terminal Bench 2.1, SWE-Atlas, and FrontierCode v1.1. The first three were conducted on the company's internal MSBench platform using GitHub Copilot CLI, while the results of the fourth were taken from the independent service Artificial Analysis.
The lightweight Code 2 Flash is designed for short queries and repetitive tasks. According to Microsoft's assessment, in three programming tests it showed comparable quality to Claude Haiku 5.5 at 20% of the estimated cost, and with GPT-6 Luna at 65%.
When comparing costs for both models, Microsoft used a promotional price with a 50% discount. The company also clarified that the developments are not yet optimized for fundamental scientific research and long-term autonomous tasks.
MAI Code 2 and Code 2 Flash are available in GitHub Copilot. The lightweight version is additionally undergoing limited preliminary testing in Microsoft Foundry.
MAI Cyber 2 Flash tested for vulnerability detection
For cybersecurity tasks, Microsoft introduced MAI Cyber 2 Flash. The model is designed for analyzing program code, detecting vulnerabilities, and preparing recommendations for their elimination.
In the CyberGym-E2E test, it successfully completed 92.3% of tasks compared to 91.5% for GPT-6 Astra. At the same time, MAI Cyber 2 Flash used half the reasoning budget.
According to Microsoft's calculations, the cost of completing one task was less than 2% of the expenses for GPT-6 Astra. This metric refers to testing under API rates and does not reflect actual client bills.
When verifying results, the company excluded 92 out of 905 tasks due to partial overlap of source data with materials used in training. The final comparison of all models was conducted on the remaining 813 tasks.
MAI Cyber 2 Flash supports a context window of 1 million tokens. Starting mid-December, Microsoft plans to begin deploying the model for select clients of the MDASH platform, intended for automated software security analysis.
MAI Image 2.7 topped the image editing ranking
Microsoft also released MAI Image 2.7 and the lightweight MAI Image 2.7 Flash for generating and editing images.
As of November 16, the main model ranked first in the Artificial Analysis rating for editing quality. Its score was 1215 Elo points compared to 1211 for GPT Image 2.6 Sunburst. The difference is small, so the models' positions in the ranking should not be interpreted as evidence of a significant advantage of one over the other.
The estimated cost of MAI Image 2.7 is about $35 per 1000 images, while the Flash version is approximately $10. Actual expenses depend on generation parameters and access conditions.
Along with the models, the company introduced Canvas – a workspace for creating and modifying images based on text requests and sketches. Users can edit individual elements without full re-generation.
Both models are available in public preliminary testing mode through Microsoft Foundry.
Microsoft expanded its range of voice and corporate AI services
Among other novelties is MAI Realtime 1 for real-time voice interaction. The initial cost is $0.01 per minute of audio, with additional computing usage billed separately.
Microsoft also announced the application of the speech recognition model MAI Transcribe 2 in the medical service Dragon Copilot. According to the company, the service supports about 10 million patient interactions monthly.
Another announcement was Decision 1.1 – a system for selecting an AI model for a specific task considering its capabilities and cost. Additionally, the MAI Licensed Weights program will allow individual corporate clients to customize Microsoft models and use them within the Azure infrastructure.
This material is prepared solely for informational purposes and does not constitute financial advice or a recommendation.




