Microsoft announced MAI Cyber 2 Flash on November 17, 2026, as a specialist model for security investigation and defensive work. Eligible customers will begin receiving it through Microsoft security products from mid-December, rather than as a generally available standalone release today.
What changes from the earlier model?
The announcement describes a one-million-token context window, four times the size of MAI Cyber 1.1 Flash’s. Microsoft says the larger window can support investigations involving more of a codebase and that the new model will replace the existing MAI profile for eligible MDASH customers.
The company positions the model for finding vulnerabilities, recommending fixes and investigating threats. These are product capabilities and vendor evaluation claims; Franklin has not independently tested the system on a production environment.
What do the benchmark results show?
Microsoft reports 92.3% task success for MAI Cyber 2 Flash against 91.5% for GPT-6 Astra on its CyberGym-E2E comparison. The models used different reasoning budgets, and the announcement ties the reported cost advantage to that offline evaluation.
Microsoft also discloses overlap between training sources and 92 of the benchmark’s 905 tasks. It removed those tasks for every model and reported results on the remaining 813. Readers should preserve that qualification rather than treat the figures as an unqualified result on the original full benchmark.
The listed per-token prices explain the company’s cost calculation. Microsoft explicitly says those comparisons represent offline evaluation cost, rather than actual customer spending.
What should security teams check next?
Access is tied to Microsoft security products, including Codename MDASH, and the rollout starts with eligible customers. The announcement does not establish immediate access for every organization.
Franklin analysis: confirm eligibility and the replacement schedule before planning an evaluation. Compare complete investigations, including the quality of suggested fixes and the amount of human review, rather than choosing a security model from a cost ratio alone.
