Microsoft Builds a Cheaper Rival to Anthropic's Mythos

Microsoft is preparing to launch an AI security platform called Project Perception, a tool that hunts for software vulnerabilities and helps fix them. The interesting part is not what it does. It is how the company plans to pay for it.

According to TechRepublic, which reported the plans on 17 July, Perception draws on models from three different providers: Microsoft's own, OpenAI's, and Anthropic's. A router sits in front of them and picks whichever model suits the task at hand. Routine scanning goes to something small and cheap. The genuinely hard problems get escalated.

That design is a direct answer to the economics of Anthropic's Mythos, the security-focused model that has set the pace in automated vulnerability discovery this year. Mythos is capable and expensive to run. Microsoft's bet is that most security work does not need a frontier model on every call, and that a business willing to accept slightly less capability on the easy ninety percent will save a great deal of money.

What Microsoft has not said

Quite a lot, as it happens. The company has not publicly announced Perception, confirmed a launch date, published pricing, or said which customers qualify. Reports point to a debut as early as this month, but everything in circulation so far comes from people briefed on the plans rather than from Microsoft itself. NewsBytes and Neowin both carried the same account. Treat the specifics as provisional until Redmond confirms them.

The push sits under Hayete Gallot, who took over Microsoft's security organisation earlier this year. It follows the company's multi-model agentic security system, announced in May, which used a similar spread-the-work approach and which Microsoft said topped an industry benchmark.

Buying models from your competitor

There is something worth pausing on here. Microsoft intends to sell a product that competes with Anthropic by, in part, running Anthropic's models inside it. That is not a contradiction so much as a description of where the industry has landed. Model access is becoming a commodity input, and the margin is moving to whoever assembles those inputs into something a business will actually deploy.

Anthropic gets paid either way. But it also loses the direct customer relationship, and with it the ability to set the price of the finished product. If routing works as well as Microsoft hopes, the lesson other buyers take away is that you rarely need the best model. You need a system that knows when you do.

The claim to watch is the cost one. Vulnerability discovery is a domain where a missed flaw is expensive in a way that a slightly slower scan is not, so a router that sends a subtle bug to a cheap model has not saved anyone money. Whether Perception's routing is good enough to make that trade safely is the only question that matters, and no benchmark has answered it yet.