Microsoft is preparing an AI security product, codenamed Project Perception, that deploys inside a company’s IT systems to find software vulnerabilities and fix them, The Information reported Friday — with a launch possible before July is out. That is the same job Anthropic’s Mythos does today, at prices that have made “Mythos-class” shorthand for the most expensive AI on the market.
The twist is under the hood. Project Perception doesn’t run on a single Microsoft model. It routes each security task across a mix of models from OpenAI, Anthropic, and Microsoft itself — picking whichever handles that job cheapest. So the pitch isn’t better AI than Mythos; it’s Mythos-adjacent results at a fraction of the bill. TechRepublic cites estimates putting Mythos API pricing at roughly double Anthropic’s own Opus and 82% above GPT — already two of the priciest models publicly available.
It also lands in an increasingly open feud between Microsoft and its partners-turned-rivals. CEO Satya Nadella warned on July 13 that rival model-makers were “duping” customers by using their data to improve their models, Microsoft has restricted Azure customers’ access to Anthropic’s Fable 5, and AI chief Mustafa Suleyman says the company wants to ship its own frontier models next year.
The router is the product
If a router can break elite security work into tasks that cheaper models handle, frontier capability stops being a product and becomes an ingredient — one Microsoft buys wholesale from competitors and resells inside a package it owns. Anthropic still collects API revenue on every routed query, useful for a lab negotiating $10 billion compute leases. But there is a difference between selling the ingredient and owning the customer, and Microsoft’s Windows-and-Office install base means it starts inside nearly every enterprise it wants to sell to.
Geography sharpens the edge. Anthropic’s most capable security models face US export friction — Fable 5 was blocked from shipping outside the country even as US agencies kept using Mythos and European governments and banks lobbied Washington for access. Microsoft is unlikely to hit the same wall, TechRepublic notes. If that holds, Microsoft could be selling AI vulnerability-hunting to the rest of the world before the maker of the best model is allowed to.
Our take: This is the first clean test of whether the enterprise buys capability or distribution. Anthropic’s position is that the best security model earns premium pricing; Microsoft’s counter is good-enough, cheaper, and already installed. The uncomfortable part for Anthropic: it can win every benchmark and still lose the customer relationship — while its own models power the product that beats it. Note the pattern, too. When Microsoft wants a market commoditized, it ships the commoditizer.
What to watch
- Packaging. Standalone product or bundled into Defender and E5 licensing — bundling is how Microsoft has historically settled security-market fights.
- Anthropic’s response. Whether Mythos pricing or packaging moves once a cheaper multi-model alternative is actually shipping.
- The export split. If Mythos-class tools stay largely domestic while Project Perception ships globally, the security-AI market divides on geography rather than capability — one more fault line after the labs’ split over state AI law.