Microsoft's MAI-Cyber-1-Flash beats GPT on security — but still calls OpenAI for the hard stuff
microsoft's maicyber1flash scores 96% on the cybergym benchmark, beating frontier models while cutting costs by routing 90% of tasks locally and sending only tough cases to gpt5.4.
microsoft introduced maicyber1flash, a compact security model built on its maithinking1 line, achieving a 96% score on the cybergym benchmark to detect code vulnerabilities. by resolving 90% of incidents internally and delegating only complex reasoning to gpt5.4, the system reduces costs by 50% while maintaining high performance. additionally, microsoft launched perception, an agentbased system leveraging 100 trillion daily security signals to mitigate threats in real time. this highlights microsoft's evolving strategy from an exclusive openai distributor to an advocate for open weights and model orchestration.