Microsoft is quietly building a security tool that takes direct aim at the most expensive AI in the vulnerability-hunting business. Reported first by The Information and detailed by TechRepublic on July 17, Project Perception is a multi-model platform designed to scan enterprise software for exploitable weaknesses — and to do it for a fraction of what rivals charge.
The pitch is simple: match the capability of Anthropic's Claude Mythos, the current benchmark for AI-driven vulnerability discovery, without Mythos's punishing price tag.
What Project Perception actually does
Project Perception is meant to live inside an organization's IT environment, continuously scanning code and infrastructure for security flaws, then proposing fixes. It is not fully autonomous — security teams still test and approve every change before it ships.
That human-in-the-loop design matters. AI bug finders are powerful, but a false fix pushed to production can be worse than the bug it was meant to close.
What makes Perception different is what happens under the hood. Rather than route every task to one frontier model, it uses a model-selection layer that spreads work across models from Microsoft, OpenAI, and Anthropic. Cheap, distilled models handle high-volume scanning passes, and expensive frontier calls are reserved for the few steps where they actually change the outcome.
The architecture is a cost argument as much as a technical one: make continuous, enterprise-grade vulnerability discovery cheap enough to run nonstop.
That "run nonstop" framing is the real ambition. Most security scanning today is periodic because frontier-model analysis is too expensive to leave running against an entire codebase. If Perception's routing genuinely drops the per-scan cost, the value isn't just a lower invoice — it's the ability to shift from occasional audits to constant, background monitoring that catches flaws as code changes rather than weeks later.
The Mythos price problem
Anthropic's Mythos is widely regarded as the most capable vulnerability-hunting model available. It is also brutally expensive.
According to figures cited by TechRepublic, Mythos carries an estimated API cost roughly 100% higher than Claude Opus and 82% higher than GPT — two of the priciest models on the market. For a tool you want running around the clock across a large codebase, that math gets ugly fast.
Here is the trade-off Microsoft is betting on:
| Approach | Model strategy | Cost profile |
|---|---|---|
| Anthropic Mythos | Single frontier model, always on | Highest capability, highest cost |
| Project Perception | Task-routed multi-model mix | Near-frontier capability, far lower cost |
Routing each query to the cheapest model that can handle it is the same playbook tech companies are running everywhere right now — including quietly adopting cheaper open-source models for the day-to-day work that doesn't need a frontier brain.
Why Microsoft, and why now
Project Perception fits a larger shift. Under AI chief Mustafa Suleyman, Microsoft has refocused its AI efforts on enterprise customers and on building frontier models in-house rather than distilling from partners.
That refocus has sharpened the company's rhetoric against those very partners. CEO Satya Nadella recently warned that some AI vendors were effectively using customer data to improve their own models, while positioning Microsoft as the safer, better-governed choice.
The security angle plays to Microsoft's strengths. It already sits inside most enterprises through Windows and Office, and it has shipped experimental security AI before — including Project Ire, a system that autonomously reverse-engineers malware. A vulnerability hunter is a natural next step for a company that spends every Patch Tuesday closing holes in its own products.
There is also a geopolitical wrinkle. Anthropic's export troubles — its Fable 5 model was temporarily banned from export outside the US — created friction for customers in Europe and beyond. A Microsoft security product may face fewer of those hurdles, giving it a path into markets where Mythos access is complicated.
The catch: this is still a leak
A dose of skepticism is warranted. Microsoft has not confirmed Project Perception's availability, pricing, performance, customer eligibility, or launch timing. Everything public so far comes from reporting, not an official announcement.
The tool is expected to debut before the end of July 2026, but "expected" is doing a lot of work in that sentence. Capability claims are unproven until independent testing exists, and "cheaper than Mythos" means little without real numbers on detection rates and false positives. A bug finder that misses real flaws — or floods a team with false alarms — is cheap in the worst possible way.
The Bottom Line
Project Perception is the clearest sign yet that AI security is turning into a price war, not just a capability race. If Microsoft can deliver near-Mythos results at a fraction of the cost — and if the multi-model routing holds up under real workloads — it could make always-on vulnerability hunting affordable for far more organizations than can justify it today. For now, treat it as a very credible leak with a compelling thesis, and wait for the benchmarks before believing the pitch.
