Microsoft's In-House AI Model Surpasses OpenAI at Half the Cost
Microsoft has developed an in-house AI model called MAI-Cyber-1-Flash that outperforms OpenAI's GPT-5.4 at half the cost, according to a recent preview of Project Perception, Microsoft's new AI security system.
The system uses three teams of AI agents working in loops: red agents probe for vulnerabilities, blue agents investigate and prioritize risks, and green agents write and deploy fixes with human approval.
Until now, Microsoft's MDASH vulnerability-finding harness relied on rented intelligence from OpenAI. However, the new release replaces most of that capacity with MAI-Cyber-1-Flash, a code-heavy specialist trained on Microsoft's own record of real exploits and remediations.
The model boasts a score of 95.95 percent on CyberGym, surpassing Anthropic's Mythos by about 12 points. While the score belongs to the full MDASH harness, not the small model alone, it demonstrates the effectiveness of Microsoft's in-house approach.
What's more significant is that this development showcases how platform owners can control the routing of tasks and capture the savings from dividing work efficiently. In this case, Microsoft wins on different terms: cheaper, faster, and trained on proprietary data.