Microsoft's new cyber model, 10x smaller than GPT-5.4, handles 90% of queries on its own

Microsoft's first in-house cybersecurity model will ship in Microsoft Defender on Aug. 3, the security software built into Windows and Microsoft 365. The model handles roughly 90% of queries on its own and defers the hardest 10% to OpenAI's GPT-5.4, which Microsoft AI CEO Mustafa Suleyman described as about 10x larger.

The system, called Project Perception, pairs the new MAI-Cyber-1-Flash with GPT-5.4 and Microsoft's MDASH multi-model agentic framework. Per Microsoft benchmarks released Monday, the combination scored 96% on CyberGym, a 12-point lead over Anthropic's Claude Mythos 5. SiliconAngle pegged Microsoft's exact figure at 95.95%.

Axios confirmed the 90/10 split, citing Microsoft security lead David Weston. Microsoft also claims roughly 50% lower cost than the current MDASH setup, billed by agent run time in Security Compute Units (SCUs).

The launch lands the same week Anthropic walked back Claude Mythos 5 after a US government-flagged jailbreak. Both the benchmark and the cost figure are Microsoft-supplied; CyberGym is one benchmark and is not independently replicated in the available reporting.