Google on Tuesday launched Gemini 3.5 Flash Cyber, an AI security model built specifically for finding and patching software vulnerabilities. The model is built on the Gemini 3.5 Flash foundation and positioned as a lower-cost alternative to larger AI security systems such as Anthropic's Mythos. Google describes it as a "cost-efficient and highly capable alternative" to those more expensive models (The Verge).

Gemini 3.5 Flash Cyber will be available first to governments and trusted partners through CodeMender, Google's security-focused coding agent. CodeMender can invoke the model multiple times at high speed and low cost, which Google says allows AI agents to scan more code paths and surface vulnerabilities that might go undetected in a single-pass approach (The Verge).

The multi-call strategy is central to the model's design. Google reports that 3.5 Flash Cyber achieved "competitive performance" compared to "significantly larger models" on the CyberGym AI cybersecurity benchmark when called up to five times. Rather than competing purely on per-query capability, the approach leverages inference speed and low per-call cost to run more iterations over a given codebase, trading brute model scale for iterative depth.

Concrete results from internal testing against the V8 JavaScript Engine give the claims some grounding. Gemini 3.5 Flash Cyber identified 55 unique confirmed issues in V8. For comparison, the base Gemini 3.5 Flash model identified 47, and Anthropic's Opus 4.6 identified 36. Ten of the issues found by Flash Cyber were not discovered by any other model tested (The Verge). Google did not publish the total number of known vulnerabilities in the test set, so precision and recall against a fixed ground truth cannot be independently assessed from the available data.

The competitive framing is aimed squarely at Anthropic's Mythos lineup. Anthropic's Mythos 5, released as part of the company's Project Glasswing initiative, is priced at $10 per million input tokens and $50 per million output tokens, which is twice the cost of Claude Opus 4.8. Microsoft has already adopted Anthropic's Mythos for its security checks (The Verge). Google is positioning Flash Cyber as a model that can be called repeatedly at a fraction of that cost per token, making large-scale automated code auditing more economically viable.

The broader context here is a shift in how AI security models compete. The first generation of AI vulnerability-finding tools leaned on the largest available general-purpose models, trading inference cost for raw capability. Google's approach with Flash Cyber inverts that logic: take a mid-tier model, optimize it for security tasks, and run it many times over the same code. If a specialized smaller model called five times matches or exceeds a frontier model called once, the unit economics of automated security auditing change substantially. That matters for organizations scanning large, actively developed codebases where the total surface area is vast and per-token costs compound quickly.

There are open questions Google's announcement does not address. The CyberGym benchmark results are self-reported. The V8 test involved confirmed issues, but the false-positive rate across all calls is not disclosed. And the restricted initial rollout to governments and trusted partners limits independent verification for now.

Even so, the V8 results suggest a specialized model can find novel vulnerabilities that larger general-purpose models miss, not just more of the same ones. The ten unique issues Flash Cyber surfaced that no other model detected are the most interesting data point. If that holds up under broader testing, it points to specialization producing genuinely different discovery behavior rather than simply higher throughput on the same findings.

What this enables, if the approach scales, is continuous, deeply iterative security auditing at a cost point that large enterprises and government agencies can sustain across entire codebases rather than critical components. The trade-off Google is betting on is that iteration depth plus domain specialization can substitute for raw model size. The V8 numbers are an early indicator that the bet has legs.