Microsoft's Project Perception meters security by the unit

Microsoft shipped its first cybersecurity-specific model today and wrapped it in an agentic security system that goes to public preview next week. The benchmark numbers are the headline. The pricing model is the part that will show up on your invoice, and it's the part almost nobody is reading.

What actually happened

Microsoft introduced MAI-Cyber-1-Flash, its first model built specifically for security work, running inside MDASH — the company's multi-agent system for finding and remediating vulnerabilities. Per Help Net Security, it outscored Mythos, Gemini, and GPT on CyberGym, a benchmark that measures how well a system reasons over large codebases to find real vulnerabilities, beating Mythos by 12 points. Satya Nadella's claim for the MDASH combination: "world-class performance at 50 percent of the cost of leading models."

Alongside it, Project Perception — an agentic system of red agents that probe for paths to compromise, blue agents that investigate and triage, and green agents that remediate and harden. It's integrated into Microsoft Defender first, with plans to spread across Microsoft Security products. Public preview starts August 3, 2026. TechCrunch covered the launch the same day.

Now the line item. Perception bills on consumption, in Security Compute Units, with different agents burning SCUs at different rates depending on task intensity. Not per seat. Not per endpoint. Per unit of work the agents decide to do.

Why agentic security pricing matters for your business

A per-seat security tool has a knowable annual cost. A consumption-metered agent fleet has a cost that scales with how much your environment gives it to chew on — and a noisy environment is exactly the kind that gets one of these deployed. The month you have an incident is the month the red, blue, and green agents all run hot. That's the meter working as designed; it's also a bill arriving at the worst possible time.

None of that makes it a bad product. Half the cost of frontier models on vulnerability triage is a real number, and a small team that can't staff a security analyst gets something it genuinely could not buy before. But price it the way you'd price any metered service: find the budget cap and the alert threshold before the pilot, not after. Ask what a typical month of SCU burn looks like for an environment your size, then ask what a bad month looks like. Vendors answer the first question readily and the second one slowly.

And keep the exit visible. Perception is Microsoft-first by design — it reads Defender signals, works Microsoft workflows, and lands remediations inside Microsoft products. That's the value and the lock-in, same coin. The parts you should own regardless: your asset inventory, your patch cadence, your own log of what got changed and by which agent. Those survive a vendor switch. The agent fleet doesn't.

Sources: Help Net Security, Microsoft Security — Project Perception, TechCrunch.