Microsoft dropped a bombshell on Monday. They claim their new AI cybersecurity model, MAI-Cyber-1-Fast, destroys the competition. And it does half the price of existing solutions.
This isn’t just marketing fluff. When paired with the agentic security system MDASH and OpenAI’s GPT-5.4, the stack beats Anthropic’s Claude Mythos 5. The score gap? A solid 12 points.
Microsoft’s pitch is simple but brutal: use AI to defend against AI. They are entering the arms race with a weaponized combination. It’s called Project Perception.
Benchmark Wins: Project Perception vs. Claude Mythos 5
The numbers are posted for everyone to see. On the CyberGym benchmark, the Microsoft stack hit 96%. Anthropic’s Mythos 5 lagged at 84%.
That 12-point difference matters in a game where margins are thin.
“MAI-Cyber-1-1-1 handles about 90% of thequeries. It detects vulnerabilities. It patches them. Ships them. Proves the fix. Then it defers the hard 10% to GPT-5.4.”
Mustafa Suleyman, CEO of Microsoft AI made this clear during a briefing. The handoff is seamless. MAI-Cyber-1 does the heavy lifting. It’s fast. It’s efficient. Then GPT-5.4 takes over the complex cases. It’s ten times larger. It solves what the smaller model can’t.
Together? Better performance. Lower cost.
Pricing: Security Compute Units (SCUs) Explained
How do you pay for this? Consumption-based.
Microsoft uses a metric called Security Compute Units (SCUs). Each AI agent scenario consumes SCUs. More work equals more bills. Simple.
But here is the twist. Microsoft claims near 50% cost savings compared to the current MDASH configuration. If you are spending money on legacy security agents, this new stack is a cheaper alternative that actually works better.
Why does it cost less? Efficiency. By offloading 90% of routine queries to a specialized, smaller model (MAI-Cyber-1 ), you aren’t burning expensive tokens on easy tasks. You reserve the heavy GPT-5.4 power for the actual threats.
When Does Project Perception Launch?
Public preview starts August 3. It is built directly into Microsoft Defender.
It launches with MDASH, which rolled out in May. From there, Microsoft will slowly push it across all Security products. This isn’t a beta test in a vacuum. It is going into the production environments of enterprises globally.
Why Microsoft Thinks They Win
Suleyman pointed to history. Microsoft has stewarded government and enterprise data for decades.
“It’s that data combined with expertise… that we’ve been able to drive this combined model,” Suleyman said.
They aren’t starting from zero. They have accumulated a phenomenal amount of data. World-class cybersecurity experts built the models. Anthropic, by contrast, faced backlash. Their Claude Fable 5 launch last month was pulled. Why? The US government said the model could “jailbreak” limits. It was so potent it could “break the internet.”
Microsoft is avoiding that mess. They are playing it safer. Slower. More controlled.
Anthropic’s Mythos 5 was initially released only to select agencies and tech pros. Now Microsoft is opening the floodgates.
“In conjunction… they do so at 50%of the cost.”
Suleyman called the CyberGym result “quite a remarkable.”
Remarkable, or just expected? Given Microsoft’s resources and data moat, maybe not surprising. But the industry will be watching. How many companies will switch? Can Anthropic retake the crown?
The answer isn’t clear yet. The tech is there. The price is right. But trust? That takes longer than a patch cycle.
