Microsoft has announced MAI Cyber 2 Flash, a specialist model for cybersecurity tasks that it says combines strong benchmark results with a much lower cost per task. The company plans to bring the model into its security products, putting its performance and price claims to a practical test.
Microsoft AI Watch analysis
What happened
Microsoft says the model scored 92.3% on CyberGym-E2E, a benchmark for vulnerability discovery, compared with 91.5% for GPT-6 Astra. It says MAI Cyber 2 achieved that result at less than 2% of the cost per task. Microsoft also claims it outperformed GPT-5.6 Cyber 6 on CyberGym-E2E and WebExploitBench when both were run at the same reasoning budget.
The announcement describes the model as built on Microsoft’s security data and expertise, and says the company is bringing the capability into its security products. These are Microsoft’s reported benchmark and cost comparisons, not independent evaluations.
Why it matters
Finding vulnerabilities, recommending fixes and investigating threats can demand substantial computing resources. If Microsoft’s cost comparison holds up in real deployments, a capable model that costs less to run could make it easier to use AI for more security work, rather than reserving it for a handful of expensive tasks.
The product integration matters too. A benchmark score is a starting point; the consequential question is what the model can do inside security tools, and how its performance translates to the systems teams actually protect.
Our read
This is a credible product direction with a striking price claim, but the numbers deserve a little daylight before they become a buying decision. Microsoft has put useful specifics on the table, including named benchmarks and a comparison at the same reasoning budget. The next test is whether independent results and customer experience match the company’s pitch.
What to watch
- How MAI Cyber 2 Flash performs in Microsoft security products and what tasks it will handle.
- Whether independent testing reproduces the benchmark and cost comparisons.
- How Microsoft calculates cost per task, including the models, settings and workloads compared.
Discussion spark: For security teams, should a cheaper model with near-matching benchmark results be the default choice, or should the burden of proof stay higher for AI used to find vulnerabilities?
Sources and evidence
- Source update (17 November 2026, 18:15 UTC)
not affiliated with or endorsed by Microsoft