Малые open-source модели Cisco против GPT-5.5: поиск уязвимостей за долю цены

In July 2026, Cisco released two compact open-source AI models for cybersecurity. According to the company's own tests, they find roughly 150 times more vulnerabilities per dollar spent than large general-purpose AI agents.

What exactly did Cisco release

Cisco has open-sourced two small models tailored to a single task — finding vulnerabilities in code and systems. Unlike general-purpose giants such as OpenAI's GPT-5.5, these are narrowly specialized models: fewer parameters, cheaper to run, and openly accessible, so they can be deployed on your own infrastructure.

The move fits a broader trend: instead of one giant "do-everything" model, companies are increasingly turning to a small model built for a specific task. Cisco is applying this logic to security — an area where both the cost of mass scanning and control over data matter.

• Two separate open-source models, both for cybersecurity
• Specialization — vulnerability detection
• Claimed gain — roughly ×150 in vulnerabilities found per dollar versus large AI agents
• Benchmark cited in the announcement — OpenAI's GPT-5.5
• Open license: the models can run locally without sending code to external services

Why is it 150 times cheaper?

Model size drives the economics. A large agent built on GPT-5.5 spends expensive inference on every run, and scanning a large codebase means thousands of such runs. A small model performing the same task costs orders of magnitude less per call, and it's precisely this per-call price difference, multiplied by the volume of checks, that produces the claimed roughly 150-times gap in vulnerabilities found per dollar.

An important caveat: the ×150 figure comes from Cisco's own internal tests, not an independent benchmark. In its brief announcement, the company disclosed neither the exact measurement methodology nor the datasets used to compare the models. Independent checks on public codebases will show the real gap.

According to

Cisco's own tests, the compact models find roughly 150 times more vulnerabilities per dollar than large AI agents — as reported by The Decoder, citing the company's announcement.

What changes for security teams?

For cybersecurity teams, the key point here isn't just price but also openness. A proprietary agent requires sending code fragments to someone else's cloud, which is often unacceptable for security work: source code and discovered vulnerabilities are sensitive data. Cisco's open-source model can be deployed inside the perimeter and scan code without letting it leave.

The second shift is scan frequency. A small model can realistically run continuously in a CI/CD pipeline on every commit, whereas an expensive GPT-5.5-based agent, due to its cost, is more often run selectively. With a roughly 150-times price difference, continuously scanning every change becomes economically justified rather than an occasional luxury.

Finally, a major industry player stands behind the release. Cisco is one of the leading suppliers of networking and security hardware, and the fact that it is the one open-sourcing specialized models raises the odds that they'll be adopted in real SOC and DevSecOps processes rather than left as an experiment.

What it means

Cisco is betting that in cybersecurity, narrow specialization and open access matter more than universal power: two small models, by its own tests, are cheaper and more private than a large agent, and the price gap is measured not in percentages but in hundreds of times. There's no independent verification of these numbers yet — and that's exactly what will show whether the ×150 claim holds up against real codebases.