Some news makes you grin for a second — and this is one of them. According to a report from The Information, Microsoft is preparing an AI security tool, internal codename Project Perception. It scans enterprise environments for weaknesses and proposes fixes. The catch for Anthropic: it’s positioned directly against Claude Mythos 5 — and yet Claude itself is running under the hood.
A router, not a single model
The real trick of Project Perception is the architecture. Instead of betting on one model, a routing layer hands each task to whichever model does it best — drawn from Microsoft, OpenAI and Anthropic.
One model is stronger at reviewing source code, another spots unusual behaviour, a third writes usable fixes. Vulnerability analysis, code review, threat detection, incident response — each goes wherever the balance of quality, speed and price fits best. For the customer that means advanced AI security without committing to a single vendor.
The point of attack is price
And this is exactly where Microsoft aims at Anthropic. Claude Mythos 5 is the most visible AI system purpose-built for cybersecurity — but it’s expensive. Anthropic lists Mythos 5 at $10 per million input tokens and $50 per million output tokens, and access stays limited to a small set of vetted partners.
Microsoft’s math is simple: route queries intelligently to cheaper models, and AI-driven vulnerability hunting gets cheap enough to run essentially nonstop. Not the one brilliant scan — the permanent one. Add Microsoft’s home-field advantage in distribution — Azure, Windows, GitHub, Microsoft 365 and Defender. For a Windows-heavy enterprise already running Defender XDR, Defender for Cloud, Entra ID and GitHub, a tool that safely combines those signals can answer a question standalone scanners can’t: is that theoretical CVE actually exploitable in my specific environment?
My take
The interesting part isn’t the tool itself — it’s the move behind it. Microsoft turns the competition between models into a platform opportunity, using Anthropic’s own model as one ingredient among several to undercut Anthropic’s premium product.
For Anthropic that’s a double-edged sword. On one hand, revenue flows whenever Claude sits in the router. On the other, it erodes the very exclusivity that justifies Mythos 5’s price. If “the best cybersecurity model” ends up being just one component in Microsoft’s routing logic, then Microsoft isn’t selling the model — it’s selling the orchestration around it. And orchestration has always been the business with the bigger lever.
Project Perception is still just announced. But the direction is clear: the next round of competition won’t only be about who has the best model — it’ll be about who picks between them most cleverly.
Sources: