Microsoft MAI Cyber 2 Flash Tops GPT-6 Astra on CyberGym-E2E at Lower Cost
Microsoft AI has introduced MAI Cyber 2 Flash, a specialist cybersecurity model that the company says outperformed GPT-6 Astra on its CyberGym-E2E benchmark while costing less than 2% as much per task in offline evaluati
Microsoft AI has introduced MAI Cyber 2 Flash, a specialist cybersecurity model that the company says outperformed GPT-6 Astra on its CyberGym-E2E benchmark while costing less than 2% as much per task in offline evaluations. The result matters because it shifts attention from headline model size to the economics of completing security work, including vulnerability discovery, remediation proposals, and threat investigations.
According to Microsoft's official MAI Cyber 2 Flash announcement, the model achieved a 92.3% task-success rate on CyberGym-E2E, compared with 91.5% for GPT-6 Astra on the same tasks. Microsoft also says MAI Cyber 2 Flash used half the reasoning budget in that comparison. The company plans to begin rolling the model into Microsoft security products, including Codename MDASH, in mid-December 2026.
The announcement is not a claim that every security workload will be cheaper or better with MAI Cyber 2 Flash. Microsoft is reporting results from specific offline workloads and configurations. Customer spending and results will vary with the task, model setup, token use, and evaluation methodology. Still, the figures offer a concrete example of how specialized models may change the cost of applying AI to repeatable security tasks.
What Microsoft reported on CyberGym-E2E
CyberGym-E2E is presented by Microsoft as a real-world cybersecurity benchmark. For its official comparison, Microsoft evaluated 813 tasks. It excluded 92 of the benchmark's original 905 tasks because they came from open-source projects and vulnerability reports that were shared with training data.
That exclusion is important context. It is intended to reduce the chance that reported results reflect familiarity with examples that may have appeared in model training data. It also means the 92.3% and 91.5% figures should be understood as results on the remaining 813-task evaluation set, rather than the full original benchmark.
Microsoft's central comparison is notable for three reasons:
- Task success: MAI Cyber 2 Flash recorded 92.3%, versus 91.5% for GPT-6 Astra.
- Reasoning budget: Microsoft says its model used half the reasoning budget for the reported CyberGym-E2E comparison.
- Reported economics: Microsoft estimates that MAI Cyber 2 Flash came in below 2% of GPT-6 Astra's cost per task in its offline evaluation.
| Measure | MAI Cyber 2 Flash | GPT-6 Astra |
|---|---|---|
| CyberGym-E2E task success | 92.3% | 91.5% |
| Reasoning budget in Microsoft's comparison | Half of Astra's budget | Reference budget |
| Listed input price per 1 million tokens | $0.10 | $10 |
| Listed processing price per 1 million tokens | $0.01 | $1 |
| Listed output price per 1 million tokens | $0.50 | $50 |
The listed token prices explain Microsoft's cost framing. The company lists MAI Cyber 2 Flash at $0.10 for input, $0.01 for processing, and $0.50 for output per one million tokens. For GPT-6 Astra on Azure OpenAI, Microsoft lists $10, $1, and $50 respectively. A lower token rate does not automatically produce a lower total bill, since workloads can use very different input, processing, and output volumes. Microsoft's sub-2% result is therefore best read as an evaluation-specific efficiency finding, not a universal pricing guarantee.
Microsoft also reports favorable WebExploitBench results against GPT-5.6 Cyber at the same reasoning budget, although the supplied announcement summary does not provide a numerical result for that comparison. GPT-6 Cyber remains available to select customers, and Microsoft says it will share further comparative results in the coming weeks.
Why a specialist model could matter for security teams
MAI Cyber 2 Flash is designed for security work rather than as a general-purpose model. Microsoft says it can find vulnerabilities, suggest fixes, and investigate threats at scale. Its one-million-token context window is four times the size of MAI Cyber 1.1 Flash, giving it more room to consider broader codebases and evidence during multi-step investigations.
For businesses, the practical question is not whether a benchmark winner can replace security expertise. It is whether AI can make a limited team more effective on defined, reviewable work. A model that can assess more code or investigation context in one session, while consuming fewer tokens, could make it more feasible to use AI assistance more often instead of reserving it for exceptional cases.
Microsoft says early Perception evaluations show comparable performance to leading models for investigation-agent and security-detection authoring. Insight Enterprises reportedly found MAI Cyber models around 20% token-efficient compared with some GPT-based competitors in independent tests. These findings add context, but they are not a substitute for testing the model against an organization's own code, security data, workflows, and human review standards.
The rollout into Microsoft's security products may also matter as much as the raw benchmark result. Integrated availability can reduce the work required to move findings between a model and the tools where teams investigate, triage, or respond. Microsoft has only specified that rollout will start in mid-December 2026, so the exact product availability, feature scope, and customer access details remain to be seen.
For companies evaluating cybersecurity AI, the announcement points to a more useful buying question: what is the cost of completing a reliable task, not simply the cost of a model call? Teams should evaluate task quality, context requirements, token consumption, integration fit, and the review steps needed before acting on a model's recommendation.
Security AI can deliver value only when it fits real workflows, data sources, and review practices. Scalevise helps businesses identify practical use cases, assess model and integration options, and turn promising capabilities into dependable processes through AI workflow automation services. If your team is considering AI-assisted security investigations or remediation workflows, start by discussing an AI automation project with Scalevise.
Frequently Asked Questions
What is MAI Cyber 2 Flash?
MAI Cyber 2 Flash is Microsoft's specialist cybersecurity model. Microsoft says it is designed to find vulnerabilities, propose fixes, and investigate threats at scale.
How did MAI Cyber 2 Flash perform against GPT-6 Astra?
In Microsoft's CyberGym-E2E offline evaluation, MAI Cyber 2 Flash achieved 92.3% task success, while GPT-6 Astra achieved 91.5% on the same tasks. Microsoft says MAI Cyber 2 Flash used half the reasoning budget.
Why does Microsoft say MAI Cyber 2 Flash costs less than 2% per task?
Microsoft's offline evaluation used much lower listed token prices for MAI Cyber 2 Flash than for GPT-6 Astra on Azure OpenAI. The company says this produced a sub-2% relative cost per task on the evaluated workloads. Actual spending will vary by workload and configuration.
When will MAI Cyber 2 Flash be available in Microsoft security products?
Microsoft says rollout into its security products, including Codename MDASH, is planned to begin in mid-December 2026. The announcement does not provide fuller availability details.
Conclusion
Microsoft's MAI Cyber 2 Flash announcement is significant because it pairs a slightly higher reported CyberGym-E2E task-success rate than GPT-6 Astra with a far lower evaluation-specific cost per task. The results are not a universal measure of security AI performance or spend, but they reinforce the value of testing specialist models on the real tasks, data, and review processes that determine whether AI delivers practical security value.
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β full credit and traffic to the original publisher.